Debanna Das

Travel sites answer a query with hundreds of undifferentiated results. This agent answers it with a recommendation. A LangGraph state machine drives the conversation: an entry node classifies intent and loops back for anything missing, GPT-4o-mini parses the exchange into a Pydantic-validated query, and a smart router decides via MD5 route hash, a 30-minute cache TTL, and an LLM judgement on freshness keywords, whether to serve cached data or open a WebSocket for live flight and train availability. Geocoding resolves place names against a local database of 8,000+ Indian railway stations before falling back to the Google Maps API, so the common case never leaves the process. The routing decision is the whole design: it is what keeps latency and API spend down without ever serving a stale fare.

The assistant generated thousands of conversations a week and none of it was being read. This engine makes it legible. A three-stage pipeline ingests raw CSV logs, cleans and caches them, then derives features: temporal buckets, session boundaries from idle gaps, and language detection done by Unicode range rather than a model, Hindi and Tamil are identifiable by script, so a dependency on spaCy or TensorFlow buys nothing. On top of that sit a Random Forest that predicts reply type, K-Means segmentation over user-level features, and rolling DAU/WAU/MAU with stickiness and retention cohorts. The modelling is deliberately shallow: a product team can act on an explainable cluster and cannot act on an embedding.

At GENIE AI, I worked across the full ML development lifecycle from studying LLM architecture fundamentals (transformer attention mechanisms, positional encodings, and layer normalization strategies) to fine-tuning GPT models for domain specific applications in healthcare diagnostics and educational content generation. Developed systematic evaluation frameworks that benchmarked fine-tuned outputs against ground-truth behaviors across precision, recall, and domain specific accuracy metrics. Led the system architecture design for GenCanvas, defining the end-to-end data flow: user natural language inputs processed through spaCy NLP pipelines for entity extraction, routed to domain specific GPT endpoints with custom system prompts, validated against JSON schemas, and rendered as interactive canvas elements via real-time WebSocket connections.

Reddit closed self-service API app creation in June 2026, so PRAW is unavailable: there is no client_id left to obtain. This agent reaches Reddit through Composio’s managed OAuth app instead, which is the ordinary third-party-client model rather than a way around the policy. It runs twice a day with no server, no database and no always-on process: Actions cron is the scheduler, the runner is the compute, and a state.json committed back to git is the store, which also produces the activity that stops GitHub disabling a scheduled workflow after sixty days. Access was never the hard part though. Automating a social presence usually produces spam, and a bot that replies to N posts a day is indistinguishable from one, so the whole architecture is built around restraint. Nine cheapest-first guard clauses throw out most candidates on pure arithmetic before any model call costs anything, and the pass rate is a diagnostic rather than just an outcome: 6% with well-chosen keywords against 40% with generic ones, measured over 168 live posts. What survives is scored with relevance weighted three times above velocity, freshness and crowding, with log1p damping the last two so a single runaway thread cannot dominate the ranking. The ranking schema then carries an explicit skip_all escape hatch, so the model is allowed to answer that nothing today deserves a reply. reddit_io.py is the one hard boundary in the codebase: agent.py never sees a Composio object, a tool slug or a raw Reddit payload, only Candidate dataclasses, so if the vendor disappears it is one file to rewrite rather than the agent.

Three cooperating planes. A client plane where the entire study lives in browser memory: JSZip extraction, Cornerstone3D rendering with HU-accurate windowing, nothing persisted, so there is no PHI at rest. An API plane of Next.js route handlers that hold the keys, gate every request behind a shared secret and a fixed-window per-IP rate limiter, and route server-to-server fetches through an SSRF host:port allowlist. And an intelligence plane split in two on purpose: one providers.ts facade over three OpenAI-compatible vendors: OpenRouter, NVIDIA Build and Google Gemini, so swapping a provider is a config change rather than a code change, beside an in-process ToolRegistry of seven deterministic radiology tools driven by a ReAct loop. The registry mirrors MCP’s name/description/inputSchema/execute contract without the transport, so it ports to real MCP later. runAgent takes its chat-completion function and its registry as injected dependencies, which is why 360+ tests exercise tool dispatch, error paths, the turn budget and the trace shape with no network at all. Every tool returns evidence carrying the slice it came from, and that one field is the reason a finding in the report is clickable back to the pixels behind it.