Why Bangladesh, why now
Three things are true at once: the policy window is opening, the talent is cheap and underused, and nobody has yet built the research-layer company. Agami Labs is that company. DeepMind was not built in a vacuum; it was built where talent, capital and a national moment met. Bangladesh’s version of that moment is 2026–2030.
A national AI policy is being written right now
Bangladesh’s National AI Policy 2026–2030 is in draft: public consultation closed in February 2026 and Draft V2.0 is under review. Posts, Telecom & IT Minister Faqir Mahbub Anam announced the “human-centric” policy at WAIC Shanghai in July 2026. The draft names seven priority sectors (public services, manufacturing, agriculture, smart mobility, skills & education, finance & trade, healthcare), AI Centres of Excellence, and a sovereign AI Data Exchange. Companies that exist while the rules are written get to help shape them, and get first access to the programs that follow.
Sources: tob.news · The Business Standard · tech.bdai.app
Hi-tech parks and a $13M national AI hub
BHTPA is building toward 28 hi-tech parks (5 operational). Kaliakoir, the country’s first, has 85 companies allotted and $283M invested, with tenants including Honor, Xiaomi and Hyundai. Separately, a $13M Koica-backed national AI hub (BHTPA, 2026–2029) targets 865 AI specialists trained and 30 AI startups supported.
CAVEAT The Koica project’s approval was delayed at the Planning Commission, putting the grant at risk. Treat it as an opportunity to pursue, not a certainty to plan around.
Sources: The Business Standard · TBS, Koica hub
World-class hunger at a fifth of regional cost
A senior AI/ML engineer in Dhaka costs BDT 120,000–200,000/month (roughly $1,000–1,700); a junior, $300–550. Mid-level software engineers run $250–1,000/month against India’s $550–2,600. AI skills carry a 30–50% premium locally because supply is very low, which means the company that trains talent owns talent. BUET and the University of Dhaka anchor the research end; roughly 2,500 startups and 650,000+ freelance digital workers fill out the ecosystem.
Sources: Nucamp salary benchmarks · bdtechjobs.com
The honest constraint
Connectivity is the hard limit: average mobile internet is 9.2 Mbps against a 64.2 Mbps global average, only 44.5% of the population is online, and digital literacy sits near 8%. That kills most consumer-AI plays and barely touches this plan: labelling is exported over dedicated lines, enterprise AI is B2B, and robotics works offline at the edge. The plan is deliberately built for the Bangladesh that exists, not the one in pitch decks. win-tk.org, Feb 2026
One group, three engines, every major industry
The structure mirrors how Indian IT majors cross-subsidise research from services revenue, with one addition India’s services giants never had: a physical robotics arm fed by the same data. Each division must be independently viable; together they compound.
Data Labelling & AI Training Services
Annotation, RLHF and expert data services sold to the global AI labs and their vendor ecosystems. This is the revenue engine: it pays salaries from month one, trains hundreds of workers in AI fundamentals, and produces the proprietary Bangla and industry datasets nobody else has.
AI Applications & Fine-Tuned Models
Software products built on the data engine: Helm AI (the AI general manager), Bangla voice and customer-service agents, garment QC vision, fintech scoring and fraud modules, and fine-tuned Bangla models. This is where margins and multiples live.
Robotics & Automation
Machines assembled in Bangladesh from Shenzhen-sourced components: garment inspection systems, IoT retrofits, agri-drones, warehouse robots, and later service and humanoid platforms. Sold at home first, exported to every market with the same problems.
One group, many industries
| Industry | Labelling & data | AI applications | Robotics |
|---|---|---|---|
| Garments & textiles | Defect image datasets | QC vision, demand forecasting | Inline inspection rigs, IoT machine retrofits |
| Agriculture | Crop, pest and soil imagery | Advisory AI in Bangla, disease prediction | Spraying and mapping drones (as a service) |
| Banking & fintech | Call transcripts, KYC documents | Bangla service agents, fraud and credit scoring | Branch service robots (later phase) |
| Logistics & e-commerce | Parcel and route data | Forecasting, route optimisation | Warehouse AMRs, sorting automation |
| Healthcare | Medical imaging annotation (expert tier) | Triage and records AI | Hospital logistics and service robots |
| Government & education | Citizen-service corpora | Helpline agents, Helm AI for public offices | Training robots, lab automation |
Demand evidence is ranked in the research: Bangla customer-service AI and garment QC have the strongest buyer evidence today; logistics automation is the weakest near-term market and is sequenced last.
The cash engine: labelling for the world’s AI labs
The buyers are the platforms that feed OpenAI, Meta, Google and the frontier labs: Scale AI, Surge AI, Mercor, Labelbox, Handshake AI and their peers. Agami Labs enters as their lowest-cost quality delivery partner, then climbs from commodity annotation into the expert work the whole market is moving toward.
Source: Grand View Research · Grand View (collection & labelling)
Sources: Everest / PITON-Global 2025 benchmarks via piton-global.com; BD wage postings, Oct 2026. BD FIGURE IS AN ESTIMATE
Who we sell to
- Scale AI (~$29B valuation; runs the Outlier crowd platform). Note: since Meta’s ~$14.3B purchase of a 49% stake in June 2025, some lab customers pulled back and contractor rates were cut. A client to win, never the only client.
- Surge AI: a separate, competing data-labelling company in its own right, known for RLHF and expert-annotation work for frontier labs.
- DataAnnotation (dataannotation.tech): a distinct company, not connected to Surge AI. Generalists at $25–50/hr, credentialed experts $40–125+/hr.
- Mercor: over $2B gross annualised revenue by June 2026, built on expert talent networks.
- Labelbox (Alignerr): specialists at $90–200/hr.
- Handshake AI: buys operational datasets directly from companies, a route for our proprietary Bangla and garment data.
- Plus Toloka, iMerit, Prolific and the crowd platforms as volume filler.
Sources: aitraining.jobs · herohunt.ai 2026 vendor guide
The entry playbook
- Scale’s bar: it prefers delivery partners with 40+ employees, recurring-revenue design and anonymisation inside the client environment. Hiring to 40+ seats is therefore a Year 1 milestone, not an aspiration.
- Paid pilots first: contracts start as paid pilots on representative data slices, scored against gold sets, then scale on quality commitments, rework terms, security controls and surge capacity.
- Three doors in: (1) subcontractor/delivery partner to tier-1 vendors, (2) supplier to crowd platforms, (3) seller of proprietary domain datasets.
- Buyers split work deliberately across vendors to avoid concentration risk, which structurally favours a credible new entrant.
- Climb the value ladder: commodity labelling is being automated and generalist rates have been cut (one reported drop from ~$50 to ~$20/hr). The durable money is credentialed experts: coders, STEM graduates, medical and legal annotators, and scarce-language specialists. Bangla expertise is scarce by definition.
What one seat earns
Offshore annotation bills at $6–18/hour; a full-time seat (160 hrs/month) therefore bills roughly $960–2,880/month, with specialist RLHF work at $20–50/hr billing far more. Against a Bangladesh all-in seat cost of roughly $400–700, the spread is the business. The sketch below uses the research’s mid-case ($800 revenue, ~$500 cost per seat-month). Drag the seat count:
40 seats is the minimum credible scale for tier-1 vendor partnerships. 500 seats is a Year 3–4 operation.
Illustrative only, assembled from the sourced rates above, not a forecast. Prices are under pressure from AI-assisted labelling and synthetic data; QA layers and security audits eat margin; the defence is specialisation (Bangla, garment QC, agriculture) and the climb into expert annotation.
The home-market proof
Bangladesh’s own BPO association (BACCO) counts 430+ firms, 85,000+ jobs and roughly $850M in annual revenue, targeting 300,000 jobs and $5B exports by 2030. NOTE Sector figures conflict across sources (BACCO’s ~$850M vs EPB’s $724.6M for all ICT services); the EPB figure is the anchor used throughout this plan and the BACCO numbers are directional. Either way, the delivery workforce, the government support (tax exemptions, remittance incentives, hi-tech parks) and the export plumbing already exist. We are not building an industry from zero; we are aiming an existing one at its highest-value customer.
Helm AI, the flagship
Labelling pays the bills; products build Agami Labs. Helm AI is the flagship, and it is not a concept on a roadmap: it is a live product at helmmanager.com, billed on the live site as “The AI Operating System for Small Businesses,” with AI General Manager v2 in production. That small-business label is where the live site starts today; this plan takes Helm upmarket. Helm serves businesses of all sizes: small businesses, mid-market companies, and large enterprises. The job in Bangladesh is not to invent it. It is to prove it in the founder’s own businesses, enter through a paid auditing wedge, and localize it so deeply that a US competitor cannot follow.
“Your business is the ship. Helm is the captain.”
Helm connects every system a business already uses to forecast, monitor, and steer the work that runs it. In Helm’s own words: “Helm gives you the same kind of operator a Fortune 500 has on staff, for the price of a phone bill.” Not another dashboard the owner has to remember to check: an AI general manager that watches the whole business and acts.
- Who it serves: today, restaurants, convenience stores, auto shops, and multi-location retail — including owner-operated businesses with no operations team of their own. The plan extends the same AI general manager to mid-market companies and large enterprises that need one operating picture across dozens or hundreds of locations.
- Time to value: setup in under five minutes, no credit card required.
- Why it matters here: the product, the integrations and the customer proof already exist. Bangladesh gets a flagship on day one, not a prototype in year two.
One briefing, every morning: where the business stands, and the single next move Helm recommends. Product facts and quotes: helmmanager.com.
Helm acts, not just reports
The product’s live feed shows the general manager working autonomously, with the owner approving where judgement is needed:
- STAFFINGDetected a 22% sales spike on delivery and added extra staff to the Friday prep list.
- INVENTORYReordered a top-selling SKU, with vendor confirmation received.
- REVIEWSDrafted replies to new customer reviews, awaiting owner approval.
- LABORUpdated the labor forecast and flagged an understaffed Saturday shift, with a suggested staffer.
“Plug in once. Helm reads everything.”
Helm’s integration layer is why the general manager can see the whole business: point of sale, payments, accounting, e-commerce and payroll feed one operating picture. Strategically, this layer is the group’s most valuable asset. It is what makes the Bangladesh auditing wedge below possible, and it is the surface every future product in this plan plugs into.
12 modules across 4 categories
FINANCE
- Financials
- Accounting
- Sales Analytics
- Reports
OPERATIONS
- Inventory
- Tasks & Checklists
PEOPLE
- Employees
- Labor & Payroll
- Schedule
GROWTH
- Customers & Reviews
- Marketing
The AI General Manager is the twelfth module, sitting across all four categories: it is the layer owners, managers and executives actually talk to, and the one that turns eleven record-keeping tools into a managed business.
The Bangladesh strategy: prove it at home, localize it into a moat
A live Western product does not automatically win in Bangladesh. These five moves are how Helm becomes the group’s flagship there, and why the position compounds instead of being copied.
The founder already owns the beta customers
Helm’s target customers are restaurants, convenience stores and multi-location retail. The founder already owns exactly those businesses: Infinity Mart (convenience store), the Esso gas station, The Benue restaurant, and Smashed Patty (burger joint). Helm deploys across all four first, as the proving ground, before it is sold to anyone else.
Every integration bug is found at home, every ROI claim is measured on the founder’s own P&L, and the first case studies come with real numbers from real tills. No competitor pitching Bangladeshi businesses of any size can match a flagship that already runs its founder’s stores.
Enter through Integration Contract Auditing
The entry wedge comes from the founder’s original brief: Integration Contract Auditing. Before Helm can manage a Bangladeshi business, it has to connect to it, and that connection work is itself a paid engagement: Helm’s integration layer maps the prospect’s systems while its contract and audit intelligence reviews the agreements behind them, surfacing duplicate tools, unfavourable terms, leakages and renewals nobody is tracking.
The audit pays for the deployment, proves value before any subscription is signed, and ends with Helm installed as the operating layer. Land with the audit; expand with the general manager.
Bangladesh adaptation is the moat
Helm’s Western stack does not transfer directly. Square, Toast and Stripe must be swapped for local rails: bKash and Nagad payment data, the local POS systems Bangladeshi stores actually run, and review and customer-service handling in Bangla. That plumbing is unglamorous, slow to build, and impossible to shortcut from San Francisco.
This is the strategic point: a US competitor can ship the same dashboard into Dhaka tomorrow, but it cannot cheaply replicate bKash/Nagad data flows, local POS quirks and Bangla-language operations. The localization is the moat.
Every deployment feeds the group
Each Helm deployment streams operational data — sales, labor, inventory — back into the group. Aggregated and anonymised, that data improves the AI division’s models, sharpens forecasting for every other Helm customer, and builds the proprietary operational dataset at the heart of this plan’s moat (Section 02).
Helm is also the distribution channel for the AI division: fine-tuned Bangla models, service agents and forecasting tools ship to customers through Helm, reaching hundreds of businesses through one installed product instead of a hundred separate sales.
SaaS per location, tiered from SMB to enterprise
Helm is sold as SaaS, per location, per month, with pricing tiers that span SMB to enterprise rather than transplanted US pricing: an entry tier a single-store owner in Dhaka can say yes to without a finance department, mid-market tiers for growing chains that need roll-up reporting and multi-location controls, and enterprise tiers for large groups with custom integrations, dedicated support, and portfolio-wide Captain’s Briefings. The paid Integration Contract Audit offsets onboarding cost and shortens the sales cycle. The growth path is land-and-expand: one store, then the owner’s second and third locations, then mid-market chains and enterprise groups where one contract covers dozens or hundreds of sites and the Captain’s Briefing rolls up across all of them. Per-location pricing within each tier means revenue grows with the customer’s own expansion, and churn stays low because Helm holds the integrations, the history and the operating rhythm of the business.
Beyond Helm: the applications portfolio, ranked by evidence of demand
| # | Product line | Buyer | Why it wins |
|---|---|---|---|
| 1 | Bangla customer-service agents | Banks, telcos, MFS providers, government helplines | Deepest structural demand in the research. Most banks still run rule-based systems; a voice-first economy with 8% digital literacy needs agents that speak. Precedents: Intelsense AI in 10+ bank call centres; Hishab’s voice commerce pilots with City Bank. |
| 2 | Helm AI (AI general manager) | Businesses of all sizes: SMEs, mid-market chains, and large enterprise operators, from restaurants to retail | Already live as AI General Manager v2 with 120+ integrations and a working module suite; proven first in the founder’s own stores, entered via Integration Contract Auditing, taken upmarket from single sites to enterprise portfolios, and defended by bKash/Nagad and Bangla localization a US rival cannot shortcut. |
| 3 | Garment QC vision | Export factories | Software half of the robotics QC rig: >99% defect detection vs ~87% manual, ~38% less fabric waste, ~6-month payback (LightCastle analysis). |
| 4 | Fintech modules | Banks, MFS, lenders | Fraud detection, AML and alternative credit scoring from phone/MFS data. bKash’s retail AI products with Intelligent Machines reportedly delivered a 76% productivity lift and 15% monthly onboarding growth. |
| 5 | Agri-advisory AI | Agribusinesses, cooperatives, government programs | Disease prediction from photos, irrigation and price advice; policy tailwind from the plan to bring 27.5M farmers onto a digital platform. |
| 6 | Fine-tuned Bangla models (API) | Every developer and enterprise above | The research layer. TituLLMs (1B/3B) and BanglaByT5 prove academic feasibility; no production-grade Bangla foundation effort and no national benchmark exist. Our labelling operation generates the proprietary conversational data to build the first one that matters, on fine-tuned open models, not from-scratch pretraining. |
Demand ranking and precedents: research §7. Sources: LightCastle Partners · TituLLMs (arXiv) · BanglaByT5 (arXiv)
Robots Bangladesh needs, built in Bangladesh
The model is the one Bangladesh already runs for motorcycles and EVs: source components from Shenzhen, assemble, integrate and maintain locally. The honest economics: assembly alone is thin. The value is in the integration, the software, the customisation and the maintenance contract, which is exactly where a group that also owns the AI and the data wins.
PHASE 1 · BUILD FIRST
Garment QC vision systems
Camera rigs plus our defect models, installed inline at export factories. The wedge product: priced at BDT 50–150 lakh per system against buyer chargebacks factories already pay, with a ~6-month payback on LightCastle’s analysis. Brands (H&M, Zara, Uniqlo) are already automating 35–40% of production processes; factories that cannot prove quality lose orders. Export: every RMG country (Vietnam, Cambodia, Ethiopia, Pakistan) has the same inspection problem.
PHASE 1 · BUILD FIRST
IoT sewing-machine retrofits
Sensor kits that connect existing machines to production monitoring, at BDT 20–50 lakh per factory line-up. The precedent is proven at scale: Snowtex Group has connected ~10,000 machines since 2023 and reports up to 25% productivity gains (company-reported). Cheap, fast to install, and every connected machine streams training data to Engine 2. Export: retrofit kits ship anywhere sewing happens.
Source: textalks.com
PHASE 2 · SERVICE MODEL
Agricultural drones, sold as a service
Smallholders cannot buy drones; cooperatives and agribusinesses can buy spraying and mapping per acre. Multispectral imaging feeds the agri-advisory AI: plant health, soil moisture, disease detection. Government momentum is real (a digital platform for 27.5M farmers; Dhaka has already sought Chinese agri-drone technology), but adoption is nascent, so the service model, not hardware sales, carries the risk. Export: the same delta-farming economics apply across South and Southeast Asia.
Sources: tob.news · digibanglatech.news
Warehouse AMRs & logistics automation
Autonomous mobile robots and sorting systems for Daraz-scale e-commerce and 3PL warehouses. The technology is proven next door (Addverb, GreyOrange), but the research found no warehouse-robot deployments in Bangladesh yet and few near-term buyers. We build integration capability on imported platforms first, assemble locally once order volume justifies it, and do not burn capital ahead of the market.
Healthcare & service robots
Hospital logistics robots (medication, linen, meal delivery), reception and telepresence units. Healthcare robotics in Bangladesh is greenfield, and our route in is structural: EATL’s health-sector ties through International Medicare Ltd. and Imperial Medical College & Hospital provide pilot sites no competitor can match. UNPROVEN MARKET Pilot first, scale on evidence.
Humanoid & general-purpose platforms
The DeepMind endgame: embodied AI on assembled humanoid platforms (reference kits price at $23–25.5k from UBTECH-class suppliers today). This is a Years 5–10 research program, funded by the other engines, aimed at export markets where labour costs make the maths work years before they do at home. India’s iHub already ships service humanoids to the UAE and Saudi; the regional race has started.
Parts from China, value in Bangladesh
- Actuators are 40–60% of a robot’s bill of materials (motors, harmonic reducers, servos); Chinese suppliers undercut imports by 30–40%, and complete 6-axis industrial arms (15 kg payload) list at $7,800–8,200.
- Shenzhen/Huaqiangbei gives 3–5 day samples, 7-day custom PCB prototyping, depth cameras, LiDAR, IMUs and battery packs off the shelf; the same supply chain Tesla Optimus and Figure draw from.
- Chinese humanoid reference BOM at scale: $16,800–21,500, against $32–45k projected for US-built equivalents. Assembly location is a cost weapon.
- Bangladesh already runs this play: SKD/CKD knockdown assembly is how its motorcycle and EV industries work. Robotics slots into an existing customs and factory culture.
- OPEN QUESTION The import-duty treatment of robot parts (CKD/SKD rates) was not established in the research and is critical to the assembly maths. Verifying it with NBR/BHTPA is a first-90-days action.
- Positioning rule: sell augmentation, not replacement. Automation anxiety is a real political risk in a garment economy; our systems make factories win orders and workers move up to supervision and QC roles.
Sources: shenzhenroboticsmarket.com · valumech.com · marketplace listings via accio.com (directional)
Who we hire, in what order
The shape follows the engines: a delivery organisation that can scale to hundreds, a product team that stays small and senior, and a research core that grows only as revenue allows. The EATL hub’s training pipeline (AI, ML, IoT programs, plus Salford MSc scholarships) is the talent factory underneath all three.
Founding layer
- CEO / founder: strategy, capital, government and partner relationships.
- Strategic partner: M A Mubin Khan (EATL): facilities, government relations, convening power, health-sector access.
- CTO (AI): model strategy, Helm AI platform, research agenda.
- VP Delivery (Labelling): runs the BPO floor like a factory: throughput, quality, client SLAs.
- Head of Robotics: mechatronics, Shenzhen sourcing, assembly and field service.
- Commercial lead: sells to labs, factories and banks; owns the pilot pipeline.
Data operations
- Annotators (the volume hire): recruited for English, accuracy and trainability; paid at or above BPO band with a real promotion ladder.
- QA leads & gold-set auditors: roughly 1 per 10–15 annotators; quality is the product.
- Expert annotators: doctors, engineers, coders, lawyers for the high-rate tiers; part-time and remote-friendly.
- Data engineers & security officer: anonymisation, client-environment compliance, audit readiness.
- Trainers: run the academy with EATL; every cohort is a hiring funnel.
Product, research, robotics
- ML engineers & researchers: fine-tuning, evals, Bangla NLP/ASR; a small number of lead-level hires (BDT 250–450k+/mo) anchor quality.
- Product engineers: Helm AI platform, agents, integrations.
- Computer-vision engineers: shared between QC products and robotics.
- Mechatronics & embedded engineers: assembly, integration, field deployment.
- Field service technicians: the maintenance-contract workforce; hired from polytechnics, trained in-house.
What the talent costs (Dhaka, 2026 benchmarks)
| Level | AI/ML engineer (BDT/mo) | Data scientist | MLOps / AI researcher | AI/ML in USD, approx. |
|---|---|---|---|---|
| Junior (0–2 yrs) | 35,000–65,000 | 30,000–60,000 | 40,000–70,000 | $300–550 |
| Mid (3–5 yrs) | 70,000–110,000 | 65,000–100,000 | 80,000–130,000 | $600–950 |
| Senior (5–8 yrs) | 120,000–200,000 | 110,000–180,000 | 140,000–250,000 | $1,000–1,700 |
| Lead / Principal | 250,000–450,000+ | 220,000–400,000+ | 300,000–550,000+ | $2,100–3,800 |
Source: Nucamp AI salary benchmarks (Oct 2026 crawl), at roughly BDT 118/USD. AI roles pay a 30–50% skill premium locally; the binding constraint is senior supply, so the plan over-invests in training juniors into mids.
Year 1, Year 3, Year 10
Sequencing is the strategy: labelling first because it pays, AI products second because they compound, robotics third because it is capital-hungry. Revenue and headcount figures below are plan targets set by this plan, not market data.
Stand up the cash engine
- Incorporate at EATL Innovation Hub, Kaliakoir; secure hi-tech park incentives and space.
- Hire past the 40-seat threshold tier-1 vendors require; stand up QA, gold-set and security processes.
- Win 2–3 paid pilots (subcontracting and crowd-platform routes), convert one to a recurring contract.
- File the first grant applications: iDEA pre-seed, Startup Bangladesh Ltd, Smart Bangladesh Accelerator cohort.
- Verify the two open questions that gate later phases: robot-parts import duties, Huawei Spark enrolment.
First products in paying hands
- Helm AI private beta with 10–20 businesses across the size spectrum recruited through EATL’s network: SMEs alongside mid-market chains and enterprise prospects; integration and contract-audit engagements as the onboarding motion.
- First garment QC vision pilot installed in a live factory; IoT retrofit kits in production.
- Apply to the Core42 Partner Program (G42) and open the Hub71 conversation.
- Bangla dataset v1 assembled from labelling operations; first fine-tuned Bangla service agent in a client pilot.
From vendor to product company
- Labelling scales to 300–500 seats with a deliberate mix shift: expert annotation (STEM, coding, medical, Bangla) becomes the majority of revenue.
- Helm AI general availability: sector packs for garments, trading/distribution and services; tiered pricing from SMB to enterprise, per location per month, sold through partners and direct to large groups.
- Bangla voice agents in production at banks/telcos; fine-tuned Bangla models offered as an API.
- Robotics assembly line opens at Kaliakoir: QC rigs and retrofit kits in volume; agri-drone service covering its first districts; first export orders (QC systems to regional RMG countries).
- G42 relationship deepens: Core42 deployment for sovereign-cloud clients; Inception-style JV explored once Helm has enterprise traction.
- Raise growth capital against product revenue: regional VCs, strategic corporates, World Bank/ADB procurement contracts.
The flywheel is real
- Every division covers its own costs; the group funds research from operations, not from investors.
- Proprietary datasets (Bangla conversation, garment defects, crop imagery) are licensable assets in their own right (Handshake AI-style dataset sales).
- Agami Labs is the default Bangladeshi partner named in National AI Policy programs: Centres of Excellence, the AI Data Exchange, the Koica hub’s startup cohort.
The DeepMind of the delta
- Research lab: a standing Bangla and low-resource-language research program publishing benchmarks and models; the national Bangla benchmark is ours.
- Sovereign AI position: the private partner for government AI infrastructure under the National AI Policy, with G42/Core42-class partnerships localised in Bangladesh.
- Robotics at scale: assembly plant exporting QC systems, drones and service robots across South Asia, Africa and the Gulf; humanoid program in commercial pilots.
- Regional footprint: delivery and sales offices in the Gulf (Hub71, Abu Dhabi) and one Asian market; Bangladesh remains HQ, R&D and manufacturing.
- Capital path: growth rounds toward the multi-billion-dollar goal; a Gulf or regional listing is an option, not a need, once the group self-funds.
Moats, not market closure
- No company gets a literal monopoly; what it can get is a position competitors cannot economically attack:
- Data: the largest proprietary Bangla, garment and agriculture datasets in existence, refreshed daily by operations.
- Distribution: EATL’s government and industry network, Huawei’s 800+ enterprise clients, factory relationships across the RMG belt.
- Talent: the academy pipeline plus the research brand; the best Bangladeshi AI graduates’ first choice.
- Policy: a seat at the table while the National AI Policy’s institutions (Data Exchange, Centres of Excellence) are designed.
Huawei at home, G42 for sovereignty and scale
Two different partners for two different jobs. Huawei is already inside Bangladesh’s networks and enterprises. G42 is the Gulf’s sovereign-AI champion, and the door for a startup is a specific, named program, not a cold email.
25+ years in Bangladesh, 800+ enterprise clients
- Core network supplier to all major telcos; technical partner for Teletalk’s first commercial 5G; launched its Xinghe Intelligent Network portfolio in Dhaka in May 2025 for finance, education, manufacturing and government.
- Spark Programme (APAC): cloud credits up to $125,000, an AI development fund up to $100,000, training and go-to-market support, plus marketplace distribution.
- What we want: cloud credits and AI tooling, the talent-certification pipeline for our academy, and introductions across its 800+ local enterprise clients.
VERIFY FIRST Spark is a 2021-era announcement, Bangladesh is not a listed Spark hub, and no Bangladesh-local Huawei Cloud region was found (nearest: Singapore/Hong Kong). Confirm current enrolment directly with Huawei Bangladesh before counting any credits. US entity-list restrictions on chip supply are a background risk to flag in any joint bid.
Sources: The Business Standard · huawei.com
The ladder, rung by rung
- Rung 1, Core42 Partner Program (launched 5 Oct 2026): explicitly open to startups and ISVs; architecture reviews, sandboxes, token credits, solution architects, joint marketing, Azure Marketplace distribution, and co-investment for priority projects.
- Rung 2, Hub71 (Abu Dhabi): the Gulf foothold and investor network.
- Rung 3, JV with Inception (G42’s product arm): the Kore.ai co-development deal (Sept 2026) is the template; it requires an enterprise product with traction, which is what Helm AI is for.
- Rung 4, MGX capital: the $100B-target AI fund writes growth cheques only at scale.
- The precedent for the endgame: in Feb 2026 G42 formalised the Condor Galaxy India 8-exaflop supercomputer with C-DAC under the India AI Mission, serving researchers and startups. A Bangladesh sovereign-compute conversation follows the same government-to-government shape.
NO BD PRECEDENT The research found no G42 engagement in Bangladesh as of October 2026. This path is inferred from G42’s published programs; rungs 1–2 are open to us today, rungs 3–4 are earned.
Sources: uaenews247.com · DataCenterDynamics · Gulf Business
Working out of Kaliakoir, with M A Mubin Khan
EATL Innovation Hub describes itself as Bangladesh’s first private-sector innovation hub, based at Block 4A, Plot 1, Kaliakoir Hi-Tech Park, Gazipur, with a Dhaka office at Rupayan Shopping Square, Bashundhara R/A. It runs software development (healthcare, education, finance), ICT training in AI/ML/IoT, university R&D links, and startup incubation, and it convenes: British and Chinese investor delegations hosted in June 2026, and the ICDSAIA international AI conference co-organised with the University of Salford (which also funds part-funded MSc scholarships, September 2026 intake).
M A Mubin Khan, the hub’s founder and MD, is also MD of EATL, a former BASIS Executive Council Director and BASIS Incubator chair, a SAARC Chamber ICT Council Director, and MD of International Medicare Ltd. and Imperial Medical College & Hospital. For Agami Labs that means: hi-tech-park facilities and incentives, a training pipeline for labelling and QA talent, government-relations credibility, investor access, and healthcare pilot sites. NOTE Khan’s bio is largely self-published and the commercial terms of any collaboration are not public; both get verified in diligence, as with any founding partnership.
Sources: eihlbd.com · mamubinkhan.com · The Business Standard
The $50M question, answered honestly
The ambition on the table is $50M of grant and program funding. The research verdict: no single grant anywhere near that size exists for a Bangladeshi AI startup. What exists is a stack. Assembled in the right order, grants de-risk the start, public funds seed the company, labelling revenue funds the middle, and procurement plus growth capital fund the scale. Here is every real component, with real sizes.
| Source | Size | What it actually is | Role in the stack |
|---|---|---|---|
| iDEA Project (ICT Division) | Up to BDT 10 lakh (~$8.5K) | Non-equity, non-repayable pre-seed grant for prototype/MVP; plus BDT 50K SME digital-adoption grants | First money; credibility stamp |
| Bangabandhu Innovation Grant | $2.57M total | Spread across 385 startups cumulatively | Early grant layer |
| Smart Bangladesh Accelerator (BHTPA) | No-cost | 4–6 month cohort, National Demo Day | Network and investor access |
| Koica AI Hub (BHTPA) | $13M program | National hub targeting 30 AI startups; approval delayed, grant at risk | Program to join, not cash to us |
| Startup Bangladesh Ltd (ICT Division VC) | BDT 500 crore fund (~$42–65M) | Equity/convertible debt, seed to growth; ticket sizes and deployment pace need diligence | The anchor local investor |
| Huawei Spark | Up to $125K credits + $100K fund | Cloud credits and AI development fund (status to verify) | Compute subsidy |
| Core42 Partner Program (G42) | Credits + co-investment | Cloud/token credits, GTM support, co-investment for priority projects | Gulf platform and credibility |
| World Bank Digital Service Transformation (P511786) | $505M program | Rural broadband, digital public infrastructure, digital skills: a procurement channel, not a startup grant | Government contract revenue |
| World Bank SITA / EDGE; ADB | $250M / ~$235–295M / $600M | AI-enabled analytics, innovation grant lines, digital-economy technical assistance | Contract and TA revenue |
| Regional & Gulf capital | $500K–3M typical early cheques | Regional VCs; MGX at growth stage. Benchmark: Indian AI startups raised $1,067M in H1 2026, but ~90% of Q3 value sat in 4 deals | Series A and beyond |
What this means for the $50M goal
Treat $50M as a cumulative capital target across the first five years (grants + credits + equity + procurement revenue), not a grant application. The sequence that gets there: iDEA/BIG grants and Startup Bangladesh seed in Year 1; labelling revenue and Spark/Core42 credits carrying Years 1–2; World Bank/ADB procurement and regional VC in Years 2–3; growth capital (MGX-class) once product revenue justifies it. Any plan that assumes a $50M grant cheque is a plan that fails in month one. Funding research §5; negative finding: no $10M+ AI grant program for BD startups found.
Crowded at the surface, empty at the research layer
Three different battlefields. In none of them does anyone yet hold the position Agami Labs is built to take: the integrated data → models → machines company.
Services firms & point startups
- IT majors with AI practices: Brain Station 23, BJIT, DataSoft, Kaz Software, Synesis IT, Selise, eGeneration, TigerIT (biometrics/national ID). Strong delivery, no research ambition.
- AI-native startups: Markopolo (martech; raised ~$1.5M incl. Stellaris), Intelsense AI (Bangla ASR, ~BDT 40M seed), Hishab (voice commerce), Intelligent Machines (vision), Socian (NLP), Speklar (Bangla call centre). Each owns one niche.
- Corporate AI: bKash, Grameenphone (86M subscribers; AI Factory with Cisco + NVIDIA), Pathao, Chaldal, Daraz. In-house, not for sale.
- Read: the chatbot/services layer is contested; the foundation and embodied layer is empty. That is where we build, funded by competing where it is crowded only long enough to cash-flow.
The platforms are customers, not rivals
- Tier-1: Scale AI, Surge, Mercor, Labelbox, and specialists iMerit (India), CloudFactory (Nepal), Sama (East Africa), Toloka, Prolific.
- Cautionary tale: Appen is declining (lost a ~$82.8M/yr Google contract; revenue −14% FY2024). Scale’s post-Meta-deal turbulence shows the risk of single-client dependence.
- Regional BPO rivals (India, Philippines, Vietnam) cost $1,000–1,200 per seat-month against our $400–700 estimate, but hold the relationships and compliance track records. We win on cost + Bangla specialisation, then on quality proof.
- Read: we compete for contracts against regional delivery centres, never head-on against the platforms that award them.
India is the power; Bangladesh is open
- Addverb (warehouse robots; ₹800 cr FY24-25 revenue; Reliance-backed; wheeled humanoid launched Feb 2026) and GreyOrange set the regional bar.
- ideaForge (drones, listed), iHub Robotics (~65 humanoids deployed across India/UAE/Saudi by Q1 2026), Unbox, Ati Motors, CynLr and others fill India’s bench.
- In Bangladesh: garment-automation equipment suppliers (Jack Machinery, HIKARI) are active, but the research found no indigenous robot OEM; BUET robotics is academic.
- Read: do not out-R&D India. Out-integrate locally: Bangladeshi factories, Bangladeshi service, Bangladeshi data, export the packaged systems India’s vendors won’t localise for.
What could kill this, stated plainly
Every item below comes out of the research, including the questions it could not answer. They are sequenced into the Year 1 plan as verification tasks rather than footnotes.
- Capital spread: three capex-hungry divisions can starve each other. Mitigation: strict sequencing, each division independently P&L-viable before the next scales.
- The commodity trap: AI-assisted labelling and synthetic data are compressing prices at exactly the tier we enter. Mitigation: expert annotation and proprietary datasets from day one.
- Client concentration: Scale AI’s post-Meta pullback hurt its whole supply chain. Mitigation: minimum three anchor clients across different buyer types.
- Talent ceiling: senior AI supply in Bangladesh is “very low”; the plan’s research ambitions depend on the academy working and on a few lead-level hires landing.
- Automation politics: garment automation touches millions of jobs; displacement projections are contested and advocacy-leaning, but the politics are real. Augmentation positioning is non-negotiable.
- Robot-parts import duties (CKD/SKD treatment): unknown, and it decides the assembly model’s margins.
- Huawei Spark current status for Bangladeshi startups: 2021-era program; no BD hub listed.
- Startup Bangladesh Ltd: is the BDT 500 crore actively deploying into AI, at what ticket sizes?
- Koica hub: approval delayed at the Planning Commission; is the $13M still coming?
- G42: no Bangladesh precedent exists; the partnership path is inferred from published programs.
- EATL collaboration terms: not public; Khan’s public bio is largely self-published. Normal founding-partner diligence applies.
- Sector data conflicts: BACCO vs EPB figures disagree; treat all ecosystem headline numbers as directional.
Sources
Research compiled 23 October 2026. Figures are as reported by their sources on that date; hedged items are flagged in the text where they appear.
- tob.news: Bangladesh drafting AI policy (18 Jul 2026)
- The Business Standard: human-centric National AI Policy (18 Jul 2026)
- Policy Magazine: Bangladesh’s AI moment (28 Jan 2026)
- tech.bdai.app: National AI Policy tracker
- TBS: $13M Koica-backed AI hub (13 Jun 2026)
- TBS: Kaliakair Hi-Tech Park investment (8 Aug 2026)
- Nucamp: Bangladesh startups to watch
- Nucamp: best-paid tech jobs / AI salaries, Bangladesh
- bdtechjobs.com: software engineer salaries
- Grand View Research: data labeling solution & services market
- Grand View Research: data collection & labelling market
- aitraining.jobs: Outlier (Scale AI) rates
- herohunt.ai: data annotator recruiting & vendor guide 2026
- PITON-Global: annotation outsourcing costs
- BDPolicyLab: ICT exports analysis (EPB data, 15 Jun 2026)
- UAE News 247: Core42 Partner Program launch (5 Oct 2026)
- DataCenterDynamics: G42/Cerebras India supercomputer
- Gulf Business: G42–India framework
- TBS: Huawei Xinghe launch in Bangladesh (26 May 2025)
- Huawei: $100M Cloud Spark programme
- entrepreneurs.bd: startup grants in Bangladesh
- World Bank: Digital Service Transformation P511786 (Dec 2025)
- ainvest: Indian AI startups raise $1B+ in H1 2026
- textalks: Bangladesh garment makers turn to AI & automation
- LightCastle Partners: AI in Bangladesh’s supply chains (Oct 2026)
- tob.news: drone farming takes flight (May 2026)
- Shenzhen Robotics Market: supply chain guide
- valumech: robot supply chain cost structure
- Addverb: industrial humanoids (Feb 2026)
- arXiv: TituLLMs, Bangla LLMs (2502.11187)
- arXiv: BanglaByT5 (2505.17102)
- EATL Innovation Hub: about
- M A Mubin Khan: about
- TBS: EATL hosts British & Chinese delegations (Jun 2026)