Hadiqa Basit

The company needed a way to intelligently match incoming clients to sales agents. The existing system was a hand-coded scoring policy that concentrated clients on a small set of agents and was painful to change.

I pulled production data and found it was falsifiable in places (agents gaming metrics), which ruled out supervised learning. After several failed prototypes and a deep research phase, I designed a contextual multi-armed bandit system, adapting the DiagUCB formulation into a hybrid of UCB and Thompson sampling. The system clusters clients by behavioural scores (evaluated by an LLM), runs three parallel bandits (one per cluster), and produces a final match score weighted by the client's cluster distribution.

I also built the full production service: FastAPI, async SQLAlchemy on MySQL, Celery + Redis for async jobs, New Relic for observability, and Dockerized deploys via GitHub Actions. After launch, I diagnosed that reward events were too rare for meaningful reinforcement, so I decomposed the reward signal into finer-grained events, lifting the reward event rate from 21% to 84%.

An e-commerce marketplace monetizes through ad-visibility packages. Pricing was a once-a-year, top-down exercise. With 200,000+ category × attribute combinations, that approach structurally cannot capture demand differences at the attribute level.

A model that produces price recommendations daily at the lowest meaningful unit (leaf category combined with item attributes) using live demand signals. The system is deliberately conservative: it tests small price increases at the margin, measures the net effect (revenue gained vs. churn), and stops or reverts when upside disappears.

In the pilot category, only 26 of 21,427 candidate cases received a recommended price change, a 0.12% recommendation rate. The system surfaces only high-confidence opportunities, not blanket increases.

The business needed an AI agent that could autonomously engage potential real estate investors over WhatsApp: qualify them, present relevant projects, and book consultation meetings, all without human intervention.

I built the agent end-to-end on a multi-tenant, multi-agent platform. The architecture uses LangChain's tool-use agent pattern (GPT-4o-mini) with custom tools: a real estate projects search API, a CRM conversation submission endpoint, currency conversion, and timezone-aware meeting scheduling.

I engineered bidirectional state tracking between the LLM and the server. For meeting bookings, I built a dual-safeguard system so that if the LLM skips the CRM submission, the server catches it and fires the submission anyway. I also built the multi-tenant WebSocket layer for the WhatsApp Business API with automatic token refresh and reconnect handling.

An AI transcription pipeline processed Urdu sales calls via Whisper. The project-name extraction field was only 37% accurate due to background noise, mid-call cut-offs, and Urdu pronunciation issues.

A custom spaCy NER model trained on the company's own call transcripts, wrapped as a FastAPI service. On top, I added a fuzzy matching layer fed by an uploadable list of canonical project names. The fuzzy matcher catches close matches deterministically, and the NER model handles anything it misses.

Users type natural-language queries like "cozy ramen spot" or "cheap brunch place" and the system needed to retrieve semantically relevant restaurants, not just keyword matches.

SBERT embeddings represent restaurants and queries in a shared semantic space, stored in pgvector with HNSW indexing for sub-linear approximate nearest-neighbor retrieval at scale. I built an NLP-based noun-pair extraction step that pulls salient (modifier, head) pairs from free-text queries so the embedding step prioritises semantically meaningful tokens over filler words.