I build AI systems for problems that already have victims, and I build them so a sceptical reader can check every claim.
0
shipped products
0
case studies
0+
hackathons
0
problem domains
0
PyPI packages
#0
Orchestrate
Stack / 01 of 06
Agents and LLMs
In the field
01 / For smallholder farmers
KrishiMitra
Thirteen modules in twelve Indian languages, free to use, each built around a question a farmer actually asks. A crop the district has never grown is refused whatever the model says.
Case study ↗
02 / For daily commuters
Floodcast Gurugram
Answers the one question a commuter asks, will my route flood and when, by matching live rainfall against 73 researched flood points instead of issuing a city-wide alert.
Case study ↗
03 / For pollution regulators
Vayu
Enforcement usually targets the dirtiest sensor rather than the site where action pays most. Ranks interventions by Gaussian-plume modelled return, then tests whether the order actually worked.
Case study ↗
04 / For city traffic planners
MargaDrishti
Models where Bengaluru loses road capacity and when, over 298k violations, with strictly causal lag features so nothing leaks from the future. Publishes what the data cannot answer.
Case study ↗
05 / For police stations and the state bureau
Kadi
A station officer gets a ranked case queue with a plain-language reason and a next action. The bureau reads state-wide analysis off the same graph of 59,985 records.
Case study ↗
06 / For the engineers who ship APIs
AGENTIQ
Turns a URL and a plain-English intent into executable functional and security tests, with every agent action passing through 19 schema-validated tools that are granted per host and audited.
Case study ↗Under the products
Adaptive Graph Search Suite, in C++20 over 1.7M junctions. Contraction Hierarchies reach 44x faster with 460x fewer settled nodes than Dijkstra, and every route is verified exact against a reference implementation rather than trusted on the speedup.
Bidirectional Dijkstra over the national highway grid, 30.9 ms in the tab
How it is organised
Measured against other people
#12 and a silver medal, up from #427
Three rounds, and the climb is the point: #427 and top 25% in June, then #290 with a bronze medal and top 15% in August, then #12 of 3,062 with a silver medal and top 1% in September. Every submission is public.
Cleared Levels 1 to 3
Deloitte's national hackathon, entered as a team of three, clearing all three levels.
Finalist, reached Level 2
Fifty days of work on Kadi, built end to end and deployed on Zoho Catalyst for the Karnataka State Crime Records Bureau. Shortlisted for the refined prototype phase. It is the project I rate above everything else I have built.
Finalist
Led team VaidyaMitra from idea to prototype: a privacy-first clinical assistant where every identifier is masked before it reaches the model, running on Amazon Bedrock in ten Indian languages.
The method
A model is allowed to propose. Something deterministic is what decides.
That separation is the decision repeated most across these projects, because it is what makes “why did this happen?” answerable by a person.
Scoring and pruning narrow a 35.7M-edge account graph, its edges weighted by entity rarity times measured fraud lift.
Greedy densest-subgraph peeling, carrying a proved one-half approximation bound, settles which accounts form a ring.
So the answer is checkable arithmetic rather than a model's opinion.
A language model turns a URL and a plain-English intent into executable test assertions.
A tool runs them, and what counts as a finding is fixed in advance against 8 OWASP API Top 10 families.
Because a model grading its own output is not evidence.