Mustakim Shikalgar

AI-powered personalized legal assistant (industry capstone, contract) with three distinct user modes across a distributed Django, React, FastAPI, LangGraph, and MongoDB stack. Served as Technical Architecture Lead for a 6-person cross-functional team across 7 sprints. Designed a dynamic FSPR knowledge profiling system that builds implicit user knowledge profiles through async background inference, targeting every response to each user's Most Critical Knowledge Gap. Architected a privacy-first two-layer system separating data orchestration from a stateless AI service enabling independent horizontal scaling, with Nginx reverse proxy for HTTPS termination and load balancing, Gunicorn multi-worker Django, Uvicorn-served FastAPI replicas behind a load balancer, and Celery with Redis for fault-tolerant background task queuing replacing Python threads. Rebuilt the LangGraph reasoning pipeline reducing cost by 85% to $0.0008 per query. Engineered token-aware rolling summarization for coherent long-session context retention and implemented lawyer-style intake behavior, context-aware legal document generation, parallel multi-dimension self-assessment evaluation, and a progressive course system spanning 9 legal domains.

Production-grade AI reliability platform that treats every agent invocation as a probabilistic operation requiring scoring, validation, recovery, and tracing. Seven decoupled microservices (gateway, orchestrator, reliability engine, guardrail layer, semantic memory, replay engine, and chaos injector) communicate over a NATS JetStream event bus with shared Postgres, pgvector, and Redis state. The reliability engine combines per-provider circuit breakers, confidence scoring, hallucination heuristics, and adaptive fallback routing to recover non-deterministically failing agents without human intervention. Every execution is event-sourced for deterministic replay: given a trace ID and a frozen model snapshot the system reproduces any historical run bit-for-bit. Full OpenTelemetry, Prometheus, and Grafana observability stack with a chaos engine that injects latency, malformed outputs, and provider outages on demand.

Fault-tolerant distributed KV store implementing Raft consensus for leader election and log replication across a 5-node cluster. Tunable consistency supporting CP vs AP trade-offs via configurable quorum-based reads.

Binary classification system for predicting semiconductor wafer pass/fail on a severely imbalanced dataset. Two-stage pipeline using L1 regularization for dimensionality reduction followed by Random Forest with balanced class weighting.

Semantic segmentation of road scenes for autonomous driving. Trained and compared three architectures — U-Net, SegNet, and DeepLabV3+ — on the Lyft/Udacity Carla simulator dataset (13 semantic classes). Fixed 7 critical bugs in the original codebase including broken IoU metric, mask decoding, and softmax/logits mismatch.

Knowledge graph integrating 10,000+ NamUs records using semantic web technologies. Replaced a $50/month GraphDB/Azure backend with FastAPI + RDFLib achieving sub-100ms SPARQL queries at zero infrastructure cost.

Genome assembler achieving 99.9% coverage on phiX174 (5,386 bp) using de Bruijn graphs and Eulerian cycle traversal. Implemented error correction including tip removal and bubble detection with multi-format support.

A browser word and math puzzle collection reconstructed from a 2022 Java CLI prototype I wrote in undergrad. Four game modes share one engine: Wordle (1 board), Quordle (4 boards), Sedecordle (16 boards, 21 guesses), and Nerdle (8-character math equation). Every day picks the same puzzle for everyone via a deterministic FNV-32 date hash; practice mode rolls a fresh puzzle on demand. Scoring is a JS port of the original Java compare() method, two-pass with proper repeated-letter handling. The Nerdle module ships its own tokenizer and evaluator that enforces operator precedence, exact integer division, and no-leading-zero rules. Stats, streaks, and daily-resume live in localStorage so the whole thing deploys as a static bundle.