Projects

Margin — making a language model safe enough to grade an exam

Scoring somebody's GRE or IELTS essay is a judgment call with consequences: a hallucinated band score sends a candidate into a real exam prepared against a false signal. Language models are fluent, confident, and wrong often enough that "ask nicely in the prompt" is not an engineering answer. Margin is my attempt to build a system where a probabilistic model does the hard reasoning and deterministic code decides what the learner is allowed to see.

Every response crosses a validation boundary before it renders. Eighteen normalizers in the frontend treat provider output as hostile input, the way you would treat a form post from a stranger. Scores are clamped to each exam's legal lattice, IELTS bands are recomputed in code as the weighted mean of their tasks so a headline can never contradict the numbers beneath it, and the client owns every state machine rather than trusting the model's account of where it is in a lesson. Each guard exists because I watched a real provider break that rule.

The runtime is the other half. Margin proxies OpenAI and Anthropic behind one interface with a reasoning-budget escalation ladder: a model that burns its whole token allowance thinking and returns nothing gets retried with more headroom instead of billing a credit for an answer nobody received. Truncation is caught through finish_reason, cache hits are accounted for, and every call reserves credits under a row lock, then finalizes or refunds.

On top of that sits an adaptive tutor — two-stage mastery, spaced review on a 3 → 7 → 14 → 30-day ladder, and error patterns from your last three graded essays pulling matching lessons forward.

Koronik — an agent is only as trustworthy as its refusal path

A bank chatbot that invents an account rule doesn't produce a bad answer — it produces a compliance incident. Koronik is Graaho's no-code agent platform, and its flagship deployment runs customer service for Islami Bank Bangladesh PLC across web, WhatsApp, Messenger, and Facebook comment threads, serving 5,000+ daily users. The hard engineering was never the answer path. It was everything the model is not allowed to do.

Retrieval is the contract. Agents answer only from tenant-approved content indexed at ingestion; anything generated outside that corpus is a failure mode, not a feature. Every response carries a confidence score, and below the configured threshold the agent stops and routes the customer to a hotline, branch, or human agent instead of producing something plausible. A refusal costs a click. A confident wrong number costs a customer. That discipline is what holds >95% answer accuracy at sub-500ms latency.

Orchestration sits on Amazon Bedrock and LangChain over serverless inference — Lambda for execution, DynamoDB partitioned by tenant for session state and strict isolation, KMS at rest. In multi-tenant RAG, retrieval leakage is a breach rather than a bug, so isolation is enforced at the partition key, not in the prompt. CloudWatch traces every agent decision: tool call, retrieval hit, latency, escalation. Usage meters in tokens against a per-tenant wallet, so cost stays attributable down to the agent and channel.

Sentiment scoring escalates frustrated conversations to human agents, who receive suggested replies rather than autonomous ones — the model drafts, a person sends. Every out-of-scope query is logged against the knowledge base, so refusals become the backlog of what the agent still cannot answer. That loop is human-reviewed today; nothing retrains itself yet. Next to close.

Computer vision inside the drawing, not around it

Architectural drawings arrive as PDFs. Easymator rasterizes every page at 300 DPI, and from there all measurement happens in image space: Euclidean distance across traced points, polygon area, circle radius. The geometry is not the hard part. A pixel means nothing until you know what it is worth.

So calibration is the contract. Each drawing carries width and height scale factors mapping canvas pixels to real dimensions, validated as positive before any conversion runs, and every measured length is divided by them before it reaches a material. Get that wrong, and 7,410 catalog materials are priced against a fiction.

Recognition stays human-in-the-loop by design. The contractor traces the wall; the system measures it. Automatic detection of walls and openings was the obvious next step, and we did not take it, because a misdetected edge produces a confident wrong number nobody can see. A badly traced edge is visibly wrong.

The pipeline reports on itself, recording duration, point count, and success or failure for every calculation across a rolling window. One gap stands: when calibration is missing, the converter falls back to raw pixel units instead of refusing. Fail-open, and next to fix.

AlgoRec: making a learned system dependable enough to bill for

Recommendation is a deceptively hard reliability problem. The signal is implicit — nobody rates anything; they just buy — the matrix is overwhelmingly sparse, and a degraded model fails silently, returning ten confident items that are wrong. AlgoRec serves that as metered multi-tenant SaaS, where every API call is charged. Plausible-but-wrong is not a modelling inconvenience. It is a billing dispute.

Classical ML carried most of the load. We modelled implicit feedback with ALS matrix factorization, applying confidence weighting so a repeat purchase outweighs a single glance, and we refused to accept a model on its own metrics: every candidate was scored by AUC against a popularity baseline, because a recommender that cannot beat "show the bestsellers" is an expensive way to do nothing. Already-purchased items were masked from every ranking. For e-commerce, we engineered features directly — purchase frequency, per-customer z-scored quantity, quintile price bands, and category-preference vectors from one-hot means.

Serving is a search problem. Scoring a full catalogue per request does not survive contact with production, so embeddings are precomputed during training and queried through an approximate nearest-neighbour index in 64 dimensions.

MLOps makes that repeatable. Training is dispatched over Redis to isolated Celery workers — one task per process, hard time ceilings, in-task memory guards — and every run is versioned in MLflow with its artifacts in object storage. Exactly one run is ever marked servable, and quota is checked before inference rather than after.

Two real problems, solved: predicting which restaurant a delivery customer orders from next across 18,948 users, and turning a WooCommerce catalogue into a live recommendation endpoint on a five-minute sync.

KhaoDao GPT is an advanced AI-powered chatbot designed for the food delivery industry. The project leverages Llama or Ollama large language models (LLMs) integrated with Elasticsearch to provide users with real-time, contextually accurate information about local restaurants, menus, and food items. KhaoDao GPT interprets user inputs through natural language processing, retrieves relevant data from Elasticsearch, and generates responses guiding users through the ordering process. The bot handles everything from suggesting restaurants based on location to crafting detailed order payloads, ultimately facilitating seamless interactions and order placements through the backend API. This innovative solution combines the power of LLMs with Retrieval-Augmented Generation (RAG) techniques to deliver an intelligent, responsive, and user-friendly experience.

Our project represents a groundbreaking advancement in traffic analysis and toll collection systems, leveraging cutting-edge technologies including computer vision, deep learning algorithms, and real-time data processing. Operating in the dynamic traffic environment of Bangladesh, our system is designed to enhance traffic management, ensure road safety, and streamline toll collection processes.

By harnessing advanced technologies, our project revolutionizes traffic management and toll collection processes in Bangladesh. It promotes road safety, reduces congestion, optimizes resource utilization, and enhances overall transportation efficiency. With real-time insights and automated functionalities, our system empowers authorities to make informed decisions and uphold traffic regulations effectively.