Md Jahid Hasan

01 · Stage one

PromptProof

Can I trust this output?

A self-correcting prompting engine that verifies information rather than assuming it. The demonstration task is a grounded claim checker: paste a paragraph, and the engine decomposes it into atomic claims, checks each against live web evidence, and returns Supported / Refuted / Unverifiable with a citation.

• Prompt chaining — the task decomposes into extract → search → judge
• Pydantic gate checks — every intermediate output is validated against a schema, with halt, retry, and retry-with-feedback logic
• A gate-checked ReAct tool — the web-search call is itself schema-validated with controlled retry, so a broken observation can't corrupt the chain
• Feedback loop — a rule-based evaluator reviews the finished report and loops until it passes or hits a cap

A RunTrace records every step, attempt, retry reason, token count and timing. The model client and the search transport are both injectable, so the whole engine runs against mocks for offline development.

The four mechanisms above come out of a structured study document in the Agentic-AI repo covering system and user prompts, role-based prompting, chain of thought, ReAct, prompt chaining, gate checks and feedback loops. It's the groundwork this stage was built from, not a separate project.

02 · Stage two

AgentProof

Can I trust a system that picks its own path?

A from-scratch agent runtime — no LangGraph, no CrewAI — at roughly 1,500 lines of readable Python, written specifically to expose the mechanics of an agent loop and prove they work rather than hide them behind a framework.

• State-machine core — typed state in Pydantic, single-job steps, conditional routing, and a step budget so loops can never mean forever
• Gate-checked tool use — arguments validated before the world is touched, responses validated before the model sees them; transient failures retry invisibly, permanent ones become structured errors the agent can reason about
• Flight recorder — every run writes a crash-safe JSONL trace of thoughts, tool calls, observations, gate outcomes, token counts and timing; any run reconstructs fully from its file alone
• Evaluation harness — golden datasets, rule evaluators across four dimensions (task completion, quality, tool interaction, system metrics), three lenses (final response, single step, trajectory), a bias-aware LLM-as-judge, and a CI regression gate
• Trajectory viewer — any trace becomes a terminal story or a self-contained HTML report

90+ tests and the CI eval gate run fully mocked: no API keys, no network, no excuses for not verifying behaviour. Live model and live search are injected through the same interfaces the mocks implement.

03 · Stage three

GroundProof

Does the runtime hold up on a harder problem?

Retrieval-augmented generation that knows when facts expire and pays only for the context it needs. It is built on the AgentProof runtime — state machine, tool airlock, flight recorder, eval harness — and graded by it, which is the clearest evidence that stage two generalises rather than being a one-off.

• Time-aware retrieval — every chunk carries an observed_at timestamp and every query runs "as of" a moment; conflicting facts resolve through deterministic supersedence rules where the later date wins and the older is kept as dated history, and answers must carry dated citations
• Query-aware compression — between retrieval and synthesis, sentences are scored against the question and knapsack-packed into a token budget with per-sentence source attribution preserved

The full pipeline: adaptive router (none / one-shot / multi-hop) → temporal retrieval → retrieval grader gate, where weak evidence triggers a corrective fallback of reformulate, retry, then web search → supersedence resolver → context compressor → synthesis with mandatory dated citations. It runs fully offline by default, with no API keys and the corpus committed to the repo.

01 · Detection

CarDNet

Can attention tell real damage from a reflection?

A CNN-attention car damage classification network proposing a Convolutional Attention Module: parallel channel and spatial attention sub-modules inserted after each residual block in a ResNet50 backbone. Channel attention decides what to focus on, spatial attention decides where — together they separate real damage from artefacts like reflections and shadows.

• Dataset — a curated, annotated, balanced set combining private Carsales data with the public CarDD dataset, augmented to 5,000+ images per class across six categories: dent, scratch, glass shatter, tyre flat, lamp broken, no damage
• Benchmarking — ResNet50 + CAM against VGG16 and DenseNet121 backbones, and against eight state-of-the-art attention mechanisms (SENet, SRM, Gate, GCNet, CCNet, BAM, Coordinate, SPNet), with ablation studies isolating the channel and spatial contributions

02 · Grounding

GroundingCarDD

Can a description locate the damage it describes?

Text-guided multimodal phrase grounding for car damage detection, fusing visual and textual signals to localise damage precisely from natural-language descriptions, and outperforming baselines on mAP and recall. Published in IEEE Access, vol. 12, pp. 179464–179477, 27 November 2024.

03 · Production

CarDVLM

Can it produce something a business can act on?

The PhD's flagship contribution and the system that shipped. An end-to-end vision-language model combining object detection with structured reasoning to output damage type, location and severity — evaluated on both structured accuracy and semantic quality, and reported as achieving state-of-the-art performance. It was scaled into production on AWS with Metaflow inference workflows.

04 · Evaluation

CarDamageEval

How would anyone check a claim like that?

A benchmark framework for evaluating car damage assessment with vision-language models, using a dual-layer evaluation approach that measures structured accuracy and semantic quality together. It proposes a hybrid CarDD_Score metric and validates it against a curated annotated dataset with baseline experiments. Published in the proceedings of AusDM'25, the 23rd Australasian Data Science and Machine Learning Conference.

This is the same instinct as the agentic track, arrived at from the other direction: the model and the thing that grades the model are two halves of one piece of work.