Ritesh Rajput.

My workBuilt the full stack: Docling OCR ingestion, 1,431 structure-aware chunks with 1,170 cross-links, hybrid dense + BM25 retrieval with weighted reciprocal-rank fusion, exact-identifier lanes, and local cross-encoder reranking.

Outcome39/40 recall@5 (0.975) on a corrected gold set. Evidence-constrained generation with server-side source-ID, citation, quantity, and identifier validation to stop fabricated page references.

My workDesigned the two-tier router that keeps conversation on a locally hosted Qwen3-4B for fast turns and escalates to a stronger backend LLM for slower factual answers, with a harmonizer merging both paths into one audio stream.

OutcomeQuantized inference on Apple Silicon via MLX q4 (Kyutai Moshi), model lifecycle management, health checks, and Prometheus metrics. Stub mode is the default so the whole gateway runs in CI with no GPU or API key.

My workImplemented RMSNorm, rotary embeddings, grouped-query attention via scaled dot-product attention, SwiGLU feed-forward, pre-norm blocks, tied output head, and scaled init — plus weight decay applied only to 2D parameters through explicit optimizer groups.

OutcomeYAML-driven configs for 50M/100M/300M variants, byte-level BPE tokenizer training, streaming download/clean/dedup/tokenize pipeline with hash-checked idempotency, and a seeded mixture sampler. Every stage is locked behind a test gate before it counts as done.

My workBuilt the authority contract — the planner extracts intent and filters, but a deterministic rule layer decides the source. The model can never submit raw SQL; only a validated ProductFilters schema is compiled to SQLAlchemy expressions.

OutcomeMetadata-filtered dense search plus BM25 fused with reciprocal-rank fusion, parallel retrieval branches for compound questions, and a 40-question retrieval evaluation set. 100 products, 508 category nodes, 8 locations.

My workDesigned the system around two rules — every sound is positioned by word index in the frozen chapter text rather than seconds, and measurements beat models: loudness, onset, and loop suitability come from DSP, never from a model's description.

OutcomeModels are used for judgment and always produce proposals that get validated against invariants, with deterministic fallbacks. Uncalibrated scores are explicitly marked so nobody reads them as probabilities.

My workBuilt the parser, speaker/emotion annotation pass, a season bible that acts as series memory across episodes, and multi-provider synthesis across OpenAI and ElevenLabs.

OutcomeAutomatic QC transcribes every generated line and compares it to the script by normalized word-error-rate, regenerating failures. Multi-take direction renders N seeds per line and picks the best by QC errors, then local DNSMOS perceptual score.

My workImplemented 15+ classical ML algorithms and a reverse-mode autograd engine, with Dense, Conv2D, BatchNorm, Dropout, RNN, LSTM, and multi-head self-attention layers, plus SGD, Momentum, and Adam optimizers.

OutcomePackaged, tested with pytest, and documented with runnable examples. This is where I keep the math honest.

My workArchitected a multi-service platform with cross-repository interfaces defined in OpenAPI, AsyncAPI, and JSON Schema, and isolated all model access behind a dedicated AI gateway service enforcing PII redaction before any request leaves the trust boundary.

OutcomeMedia worker for validation, thumbnails, and OCR; scheduling and payment domains handled through queue-based async workflows with atomic overlap protection on every booked session.

My workBuilt 69-keypoint pose input split across left hand, right hand, body, and face, three-layer spatial graph pose encoders per sub-pose group, temporal encoders, and prior-guided fusion gates for hand pose and RGB interaction.

OutcomeOne unified text-generation objective covers isolated recognition, continuous recognition, and translation. Runs locally with a tiny causal decoder so the repo works without downloading a large language model.

My workBuilt 28 deterministic on-page checks, a placement-weighted n-gram keyword profile, and TF-IDF competitor scoring — with an optional LLM enhancement layer supporting OpenAI, Groq, and Claude keys.

OutcomeThe deterministic audit works with no key at all; the model layer is strictly additive. Groq requests are sized under a 12K-token allowance with bounded retries on 429s and user-facing recovery guidance for every failure class.

My workWorked on the context and memory model — persistent world state, per-character knowledge boundaries, and consequence tracking so side stories inherit the history that made a character interesting.

OutcomeBuilt for the Pocket FM x HackCulture hackathon, where it finished in the Top 10.

My workBuilt the FastAPI ML service that tokenizes Python ASTs and compares submissions by TF-IDF cosine similarity, plus the Docker-based sandbox runner for isolated execution.

OutcomeReact frontend, Express API with JWT auth and role checks, Prisma/PostgreSQL model, and a full Docker Compose stack.