I am currently the AI Lead at ErisAI, where I
lead research and engineering of production-grade AI systems spanning LLMs, retrieval, agentic
workflows, and evaluation infrastructure. I oversee model design, system architecture, and
deployment for reliable real-world AI applications. Previously, I was an AI
Researcher at VOLV AI and a
Research AI Engineer at BHuman
AI.
Research Work
My work spans retrieval, agent orchestration, and evaluation
infrastructure, with a focus on building reliable AI systems that move from research into real-world
deployment.
A deterministic gate verifier orders candidate repairs into DPO preference pairs; across thirteen
training runs, task-corpus diversity, not the algorithm or schedule, turns out to control
out-of-distribution generalization.
A benchmark harness for MCP tool selection that runs eight retrievers across shared registries,
pinpointing where equal-weight fusion stops helping and finding Anthropic's managed tool search
on par with local BM25.
A GRPO training method for code-repair agents that uses PatchProof's hard-gate verifier as the
reward signal directly, replacing single-shot rewards with multi-turn, per-gate credit for denser
training signal.
A cache-liveness scheduler for lead-plus-sidekick LLM harnesses, with a cost simulator, live API
verification, and a proposal for productive idle time.
DarwinPatch is a budgeted repair-search controller for coding agents that converts failed patch
attempts into bounded evidence, routes follow-up repairs through hard verification gates, and
records auditable candidate lineage.
Taneem Ullah Jan Work done at ErisAI
[project page]
A research-first look at how Synapse combines retrieval, routing, memory,
verification, and runtime controls into a dependable applied intelligence system.
tldr: An embedding-driven approach combines audio encoders with multimodal
projectors to enable direct speech-to-text processing, achieving significant performance while
training minimal parameters through block optimization.
tldr: An interactive 3D body modeling system that allows real-time
manipulation of human body shapes through intuitive measurement sliders with immediate visual
feedback in a fully navigable 3D environment.
tldr: VMVLM enhances vision-language models by using dual visual pathways,
combining Q-Former queries with direct ViT feature injection for improved multimodal instruction
following.
tldr: A unified framework for 3D virtual try-on that transforms simple 2D images
into realistic 3D
representations, by efficiently integrating clothing with the human body in a pose-adaptive manner.
tldr: A two-stage unified audio-driven talking face generation framework, which can
render
high-fidelity,
lip-synchronized videos with improved inference speed.
tldr: Current image inpainting techniques are too heavy; this paper introduces a
Row-wise Flat Pixel
LSTM, a small hybrid model for the efficient and high-quality restoration of small images.
tldr: A robust and efficient talking face generation model with highly accurate lip
synchronization
and
full facial expressiveness with more extended audio and high-quality video resolutions.
tldr: An unsupervised one-shot talking head video generation model using neural
rendering and motion
transfer techniques with non-linear transformation to animate static images.
tldr: Innovative face-swapping model that preserves the source identity features
accurately while
seamlessly adapting target attributes applicable to images and videos.
