I'm an MSc Artificial Intelligence student at Dublin City University, supervised by Prof. Alessandra Mileo, working on robust and explainable machine learning for high-stakes vision.

The thread running through everything I do is one question:

Does a model actually work — or does it just look like it does?

My dissertation surfaces a diagnostic paradox: generative restorers can add pixel fidelity (Cold Diffusion, +12 dB PSNR) while failing to recover — and often harming — downstream diagnostic accuracy, because the "restored" images push the classifier off-distribution. Explanations also drift faster than accuracy as quality drops, meaning a model can be right for the wrong reasons. Currently under review at OMIA / MICCAI 2026.

That reliability lens shapes how I work, not just what I work on. I read a paper, then rebuild its mechanism small enough to watch it run — and I report the reproductions that failed as prominently as the ones that worked. Two of the results below are negative. They're the ones I'd want you to read.

• Generative restoration for DR screening — benchmarking CLAHE, A-ESRGAN, SwinIR+GAN, Cold Diffusion and a conditional DDPM as restorers against CNNs and a ViT-Base classifier under nine synthetic degradations (blur / exposure / noise × 3 severities) on APTOS 2019. ViTs proved substantially more robust under severe noise (~62% vs ~20% accuracy) — a gap invisible on clean benchmarks.
• Explainability under distribution shift — quantifying how Grad-CAM, SHAP and attention rollout drift as quality degrades, scored by insertion/deletion AUC for faithfulness and Spearman stability rather than eyeballed heatmaps — and finding where restoration silently breaks interpretability.
• Generative-image forensics — separating authentic photographs from diffusion/GAN output via frequency artifacts and intermediate transformer features, extending toward reference-free, generator-agnostic hallucination detection for medical images.
• Episodic memory for vision-language agents — the current track, below. 🔴 active

Episodic memory for vision-language agents. (active since July 2026 · reading → reproducing → building)

VoxSight can describe what the camera sees right now. It cannot answer "where did I leave my keys?" — because nothing it saw an hour ago survives the turn it was seen in. An agent that perceives continuously but remembers nothing has no episodic memory, only perception.

The bet: this is a memory-architecture problem before it is a model problem. Rather than fine-tune a bigger VLM and hope recall emerges from scale, I'm following the differentiable-memory literature from its origin and rebuilding each mechanism by hand.

One idea connects the whole reading list: you cannot hard-pick one memory slot, because a hard pick has no gradient — so you softly blend several. NTM blends notebook rows; RAG blends Wikipedia passages. Same trick, different scale. MemGPT then breaks the pattern in the useful direction — it gives up the gradient entirely and lets the model hard-pick at runtime by calling a function — making the real premise explicit: the bottleneck was never capacity, it was allocation.

The gap I'm aiming at: MemGPT never asks whether a memory can be trusted. Every self-edit lands with no source, no timestamp, no confidence, and no record of what it overwrote — and recursive summarisation compounds the error with no path back to the original observation. For a user who can't see, a confidently wrong memory is worse than an absent one. Provenance over episodic memory is the thread, and it's orthogonal to the paging mechanism.

Rebuilt from the paper on Apple Silicon (MPS), not cloned from a reference repo. Every number here is measured in my own notebook.

15,077-passage index · 96 SQuAD-dev questions asked open-domain · exact MIPS · both equations written from scratch in log-space.

BM25 winning is the interesting result. A corpus-size artifact, not a refutation: DPR's query tower was fine-tuned on NaturalQuestions, my questions are SQuAD, and lexical overlap is unusually strong when the haystack is 15k passages instead of 21M. Reporting it beats quietly dropping the row.

Frozen DINOv2 + gated cross-attention · RefCOCOg / COCO · 4,500 images · 3,000 steps (the paper trains 20–50k).

Steering emerges — but the sanity check is the point. Feed a prompt describing a different image and localisation must collapse toward baseline. It didn't: collapse ratio 0.625 against a 0.30 pass bar, so 62.5% of the gain survived a mismatched prompt. At this scale the model is reading the image, not the text. FAIL — and it's the one number worth reporting. The gate (tanh(α·ω), α initialised at 0) was verified to reproduce frozen DINOv2 exactly at ω=0.

Every architecture I study gets rebuilt as a hand-drawn, steppable sketch you can walk through and turn the knobs on — patch size, kernel size, the prompt you steer with, how many documents you retrieve. Enter the Lab →

On the bench: JEPA · Embedding space · VLM · NTM · MemGPT · Encoder–Decoder · Diffusion

• Towards a Robust and Explainable Pipeline for Diabetic Retinopathy Classification through Quality-Aware GenAI Image Restoration — under review, OMIA / MICCAI 2026
• Autonomous Drone Navigation — pitched at the International Conference on Engineering Research and Innovations, 2025

• Robust & Explainable AI for Diabetic Retinopathy  ·  MSc thesis A five-phase pipeline: 26,370 synthetic degraded fundus images → three DR graders (ResNet-50 / EfficientNet / ViT, ordinal focal loss, quadratic-weighted kappa) → a quantitative XAI benchmark (IG, SHAP, attention-rollout, Grad-CAM) → seven restorers (Cold Diffusion, conditional DDPM, SwinIR+GAN, …) → a quality-aware trust router that decides per image whether to enhance, which grader to trust, and which explanation to return. The central contribution is a pathology-preserving DDPM — a restorer constrained not to hallucinate lesions it cannot justify.

• VoxSight — Real-Time Accessibility Co-Pilot  ·  Cursor Hackathon → product A conversational agent that "sees" for you: point a camera, hold a button, ask by voice, get a spoken answer. Audio + frame → STT → vision LLM → TTS in a single WebSocket round-trip. Built end-to-end in one night in mock mode with zero API keys — every service behind a swappable mock/live flag — then wired to real models the next morning without touching orchestration. Since grown intent routing, persistent memory, and deterministic turn-by-turn walking navigation (a routing engine, so it costs no model calls per step). The memory gap in this system is what started VoxSight Recall above.

• AI-Generated Image Detection — Etsy Research Challenge  ·  🥈 2nd place Authentic photographs vs diffusion/GAN images on a stratified ~4,340-image marketplace dataset. Hand-crafted forensics (FFT F1 0.60, ELA F1 0.65) against learned representations — and intermediate CLIP layers beat the final embedding, 0.92 vs 0.85 F1. A 6,018-dim fusion with 5-fold CV (0.87 ± 0.01), TTA and pseudo-labeling, then an Optuna-weighted 4-model ensemble reaching ≈0.94 F1. Placed 2nd on Etsy's research leaderboard and was invited to present at Etsy HQ, Dublin. (Repository private.)

• HealEdge — Federated Learning for DR Detection with XAI  ·  🥉 3rd of 70 AdvanceHealth MedTech Hackathon, Trinity College Dublin. FedAvg-trained ResNet-50 across non-IID hospital data so raw images never leave the client, with Grad-CAM heatmaps so clinicians can verify each prediction — addressing both the privacy and the trust barrier at once.

• Cyberattack Detection from CPU Usage Logs  ·  🏅 AWS hackathon winner Isolation Forest vs Random Forest on CPU telemetry — 98% accuracy, 78% recall, recall deliberately optimised over precision, because a few false alarms beat missing the spike that matters. Deployed on SageMaker with Lambda inference and SNS real-time alerts.

• 3D Spatial Reasoning — Nomad AI Neuro-symbolic system for collision-free, explainable object placement. The LLM parses intent into constraints; a deterministic geometry engine (AABB checks, clearance/walkway rules) computes valid candidates — avoiding LLM "coordinate hallucination."

Because you don't really know a mechanism until you've written its backward pass.

• RNN + BPTT in NumPy — backpropagation through time by hand, to watch the gradient flow across timesteps
• Feed-Forward Neural Language Model — a Bengio-style neural LM from the ground up: embeddings, hidden layer, softmax
• Vanilla RNN & the Vanishing Gradient — demonstrating why long-range dependencies collapse, and what fixes them

Papers read in my own words, with notebooks where the idea needs to be felt rather than read. Browse the journal →

Currently running threads on image forensics (Durall's up-convolution spectral tell, Frank's DCT counterpart, Gragnaniello's FFT baseline, and the argument that the detection arms race is unwinnable by design) and memory architectures (NTM → RAG → MemGPT), alongside Attention Is All You Need, Generative Adversarial Nets, and the MPNN → SchNet → DimeNet → NequIP arc.

• 🥈 2nd place, Etsy Research Leaderboard — invited to present AI-image detection research at Etsy HQ, Dublin
• 🥉 3rd of 70 teams — Advanced HealthTech & MedTech Innovation Hackathon, Trinity College Dublin (2026)
• 🏅 Winner — Motivalogic Hackathon, sponsored by AWS, Dublin (2025)
• 🌍 Ambassador, Thinking About Thinking — responsible-AI community (AE Global Summit, London 2026)

• LinkedIn: atharva-kocharekar
• Email: [email protected]
• Portfolio: atharvax16.github.io/Portfolio
• Architecture Lab: the mechanisms, steppable
• Paper journal: til