KrishiBot, M0NARQ, and HAWKEYE Disease represent three distinct problems that the four main projects don't touch: agricultural last-mile delivery, industrial AI for the garment sector, and public health forecasting from degraded government data. Each is a working prototype. Each is blocked by serving cost, not engineering capacity.
Vertex AI RAG on BARC and DAE crop management bulletins. Gemini Flash responses compressed to 160 characters. A USSD tree that delivers AI to button phones on any carrier. The constraint is not the AI — it is the last mile.
At $0.001 per query and Gemini Flash pricing, a pilot serving 100,000 queries per month costs roughly $100 in AI compute. The economics of meaningful agricultural AI coverage for a pilot district is under $300 per month total.
The engineering is done. The interface is validated. What scales this is a partner — a carrier agreement, an NGO, a DAE district office. Not a rewrite.
Farmer dials USSD short code or sends SMS in Bengali or English. Gateway API on Cloud Run receives raw text.
Vertex AI RAG searches the BARC/DAE corpus — official Bangladeshi agricultural bulletins and seasonal advisories.
First Gemini Flash call synthesizes a grounded response from retrieved context, constrained to factual content from official sources.
Second Gemini Flash call compresses to under 160 characters in Bengali, preserving core actionable guidance and source citation.
Standard SMS. No data connection. No app. No smartphone. Works on every carrier in Bangladesh.
"$0.001 per query. That is what it costs to answer a farmer's question about rice blast disease with a cited response from an official DAE bulletin. The economics are solved. The interface is the work."
The garment industry is Bangladesh's largest export sector and employs over 4 million people. It is also almost entirely unserved by AI tooling designed for its actual infrastructure. M0NARQ is built for Bangladeshi factories — not for the factories that can afford Siemens.
MicroViT-Tiny-Q8 running at 28ms on a recycled Android phone. 92% F1 on fabric defect detection. $150 CAPEX versus $12,000 for proprietary optical inspection rigs. The factory doesn't care who made it — it cares about the margin.
TS2Vec + Bayesian change point detection for energy anomaly. EU CBAM compliant Parquet audit logs. Tracks energy consumption in real time and flags deviations that indicate machine fault or inefficiency.
GRU-128 transfer-learned from NASA CMAPSS to textile sewing machine heads. 5-hour Remaining Useful Life warning via a $1.60 MPU-6050 accelerometer. Prevents unplanned downtime on the production line.
M0NARQ is in active accelerator applications. The 47-day payback period and 9.2× Year 1 ROI make the commercial case unusually clear for an early-stage industrial AI system.
1,084 DGHS government PDFs extracted via PyMuPDF to build a 3-year daily case series. The data existed in fragmented form across degraded government publications. The engineering problem was not the model — it was recovering the signal from the source.
Causal discovery across 36 features: temperature, humidity, VIIRS nightlights as economic proxy, population density, lagged incidence. Prophet + XGBoost dual-track. r = 0.324 temperature-dengue lag at 14 days validated at p ≪ 0.001.
The policy window is up to 14 days for pre-emptive vector control scheduling. That is the window between the forecast and the intervention. The model delivers it consistently.
VIIRS nightlight intensity as an economic activity proxy was the unexpected high-value feature — it captures population clustering and movement patterns that correlate with dengue spread better than administrative population density.
r = 0.324 temperature-dengue lag at 14 days: the strongest single causal signal in the dataset, validated at p ≪ 0.001 against the full case series.