Hi, I'm

Vishal.

I build software systems, and the AI agents inside them.

Human in the loop. Code in the wild.

psst: the little cairn is clickable

Base camp

Who I am

Vishal crouching by a lake in a pine forest, smiling, with the sun sparkling on the water
A good day by the lake

I spent seven years shipping industrial software in C++, C# and Angular before going back to school in 2023 to study how people and models actually work together.

I'm finishing an MSc in Computer Science at RPTU Kaiserslautern, specialising in Intelligent Systems (thesis graded 1.0), and I'm back at DFKI as a research assistant. Before that I spent six months at ABB, where I owned retrieval, LLM extraction and evaluation for two agentic diagnostic tools that engineers now use every day.

The engineering years are the part I'd keep. Shipping software to people who call you when it breaks teaches you to distrust a demo that works once. Most of what I build now gets an eval harness before it gets a UI, and I'd rather have six components that each run on their own than one integration that only holds together in the demo.

My papers cover the energy cost of LLM APIs and web agents, HILL, a framework that lets a person reshape a model's latent space during training instead of only judging the output afterwards, and PACE-HAR, my first-author paper on adding new sensor domains to a model without disturbing what it already learned.

Based inKaiserslautern, Germany

Right nowResearch Assistant, DFKI

StudyingMSc Computer Science, RPTU

FocusSoftware systems · LLM agents · evaluation · human-AI interaction

Waypoints

Selected work

Five in enough detail to judge, including the parts that did not work first time.

Waypoint 01

Grid Copilot

An anomaly detector tells you something broke. This one tells you why, cites its evidence, and leaves the decision to a named engineer.

Live demo A fictional network on real load profiles, with a dashboard at every level from grid to sensor. The demo replays a recorded run of the offline mock agent. The earlier HAI benchmark dashboard is here too. Open ↗

Problem

Detection on grid telemetry stops at the alert. An operator still has to work out what happened from raw signals, equipment manuals and whatever they remember about that asset, and then somebody has to own the decision.

Approach

A streaming detector flags the anomaly. An agent investigates it, pulling evidence from the telemetry window, the equipment and protocol documentation, and memory of earlier incidents on the same asset, and writes a root-cause report where every claim is cited. It reuses Cortex for orchestration and mnemos for per-asset memory.

The newer half is a simulated transformer fleet: a digital twin from grid down to sensor, a health index and risk ranking, and alarm investigations that end in a proposal. A named engineer accepts, overrides or defers it, and every decision goes into a hash-chained log. Sensor data is graded before any agent reasons about it, so a frozen sensor is marked untrusted instead of being explained. When the same override repeats on three or more transformers, it becomes a change request that a person reviews. It never changes the shared logic quietly.

Result

On the public HAI industrial-security dataset, a multivariate autoencoder, trained on HAI's dedicated attack-free file rather than a prefix of the test file, takes point-adjusted F1 from 0.57 to 0.70 at full recall. The harness also reports the stricter per-timestep score, where the same change matters more (0.26 to 0.58), and the precision-recall curve behind that single operating point, because which point is right is a product decision about how many false alarms a maintenance lead will tolerate per caught attack. Recall holds at 87% across all 118 labelled attack intervals in HAI's five test files, not just the friendlier one.

Waypoint 02

mnemos

Persistent memory for LLM agents, including the unglamorous half: forgetting.

Recording The memory browser on its forgotten tab, showing a low-confidence fact struck through rather than silently dropped. A live version with a fictional user is at the link below.

Problem

Agent memory demos well and ages badly. Facts pile up, contradict each other, and go stale, and most systems have no story at all for removing anything.

Approach

Episodic and semantic memory, retrieval that adapts its own similarity and recency strategy per user, and a reflection engine that merges, decays and forgets stale facts, with every step written to an audit trail. Three storage backends (Postgres/pgvector, Qdrant, Neo4j) sit behind one interface. It also runs as an MCP server, so tools like Claude Code can remember and recall across sessions. I use it that way myself.

Result

A FastAPI backend, a React dashboard, and a benchmark that measures recall over simulated time rather than over one impressive session. In a with-and-without test on ten realistic questions, the share answered correctly went from 40% to 100% once memory was on, with a retrieval hit rate of 100%. It is listed in the awesome-mcp-servers collection.

Waypoint 03

AI Ledger

Most AI spend reports show what was spent. This one asks whether it paid off, and says so when the data cannot tell.

Problem

Companies pay for AI seats and tokens and are asked whether it works. An honest answer needs cost joined with outcomes, a fair comparison against people who did not adopt it, and a clear line between what was measured and what was assumed.

Approach

It joins per-employee AI cost with outcome data, compares each person with their own earlier months, and subtracts what happened to people who never adopted. Every return comes with a 95% range, and a department whose range includes zero is labelled uncertain: “review”, never “restrict”. It then simulates four budget policies on the same data. Questions go to NVIDIA Nemotron on Nebius Token Factory, as a tool-calling agent that calls the app's own maths and never sees raw rows.

Result

The demo runs entirely in the browser, with Python compiled to WebAssembly, so there is no server to keep alive. Ask a question with your own Token Factory key, or read answers recorded from Nemotron 3 Ultra. An Evidence tab compares the estimate with nine published studies, and every value says where it comes from: measured, estimated, assumed or judgement. Built for the Nebius and NVIDIA hackathon.

Waypoint 04

voice-coach

A mock interviewer and a German tutor I can talk to like a phone call, and talk over.

Problem

Practising interviews and German out loud is hard to do alone. Record-and-review tools give feedback afterwards, and the speech models I tried kept misunderstanding my accent.

Approach

A web app with no server of its own. The browser streams audio both ways to Gemini Live, so you can interrupt the AI and it stops. Each answer also goes to Whisper for an accurate transcript, and a setting lets Whisper do the listening for accented speech: the phone detects when I speak, Whisper writes it down, and the model answers from that text. A text model writes the report at the end, with scores and fixes for interviews, and corrections and new words for German. Repeat mistakes are fed back into the next session.

Result

It installs on an Android phone straight from the browser. Keys, past sessions and German mistakes stay on the phone, and the hosting is a static page. Whisper listening adds roughly half a second to a second to each reply, which is the price of being understood.

Waypoint 05

po-triage

The question is not whether the AI is accurate. It is how many lines can be committed without a person, and how often one of those is wrong.

The review screen for a German purchase-order email. The original email is on the left with one line highlighted; on the right the parsed lines, where the flagged line shows no match with four candidate products, two of them blue fittings in 8 mm and 10 mm that score the same. Commit is disabled until the line has a decision.
Review screen Two blue fittings in different sizes score the same, so the line goes to a person instead of being committed. Full size ↗

Problem

Customers send purchase orders as free-text emails, including in German, and someone types each line into the ERP. A wrong line that reaches the customer costs far more than one somebody glanced at.

Approach

Matching is a cascade: exact and normalised part codes first, then fuzzy matching, and a model only inside the genuinely ambiguous band. Each line ends in one of three outcomes: committed, sent to a review queue, or no match. There is a review screen with the email beside the parsed lines, and the commit threshold is a screen that shows the automation rate and the error rate side by side, because someone in operations owns that trade-off.

Result

On eight labelled emails, 88% of lines commit without a person and none of them are wrong. At order level it is 75%, because an order is only right if every line in it is. I report both, since the customer only ever sees the order.

Other things I've built along the way

    Field notes

    Research

    1. iWOAR 2026 · Conference · first author

      PACE-HAR: Partitioned Adapter-Based Capacity Extension for Interference-Free Cross-Dataset Wearable Models

      How to add a new sensor setup to a wearable model without disturbing what it already represents. Earlier domains are preserved exactly across five datasets, and the paper says plainly what its numbers do not support.

      Continual learning · VQ-VAE · Wearable sensing

    2. PerCom 2026 · Conference

      This Is Taking Too Long — Time as a Proxy for LLM Energy Consumption

      You cannot meter a hosted model, but you can time it. On when wall-clock latency is a usable stand-in for energy draw behind an API.

      LLM energy · Inference time · API

    3. HHAI 2025 · Conference

      HILL: Interactively Guiding Model Training Through Human Intuition

      Human in the Latent Loop. Lets a person reshape latent representations during training through a distillation-inspired objective, rather than only correcting the model's outputs afterwards. I helped run the user study behind it.

      Human-in-the-loop · Latent space · Distillation

    4. AAAI 2025 · Workshop

      Sustainable Web Agents: Benchmarking Energy Consumption

      What it actually costs, in energy and CO₂, to let an agent browse the web on your behalf.

      Web agents · CO₂ · Sustainability

    The path so far

    Experience

    1. Jul 2026 – present

      Research Assistant · DFKI

      Kaiserslautern, Germany

      • An evaluation harness that benchmarks autonomous agents on success, latency and cost.
      • Helped run the HILL user study, with 14 participants.
      • First author of PACE-HAR, on adding sensor domains to a model without changing what it learned.
    2. Jan 2026 – Jun 2026

      AI Research Intern · ABB

      Mannheim, Germany

      • Owned retrieval, LLM structured extraction and evaluation for two agentic diagnostic tools, now in daily use.
      • Unified logs from around six formats into one structured record.
      • Ran the experiment that replaced whole-document prompts with similarity search: faster at the same quality, and the team adopted it.
    3. Apr 2024 – Dec 2025

      Research Assistant · DFKI

      Kaiserslautern, Germany

      • Python tooling for human-in-the-loop ML workflows.
      • Loss functions driven by real-time human feedback signals.
      • Visualisations for model confidence and embedding structure.
      • Energy and efficiency benchmarking of LLMs and agents.
    4. Jun 2021 – Feb 2023

      Senior Engineer · Rugged Monitoring

      Hyderabad, India

      • Owned a power-grid monitoring platform from a whiteboard design to sites in India, South Korea and Canada.
      • Full-stack product work in Angular and .NET Core, with device data over Modbus, IEC 61850 and SOAP.
      • When new sites went live with empty dashboards, traced it to hand-mapped signals and built a tool that generated the mapping for the team to approve.
    5. Feb 2017 – May 2021

      Software Engineer · Qualitrol

      Ahmedabad, India

      • WPF desktop applications in C++ and C#.
      • Integrated ML modules into energy-sector systems.
      • Scrum Master for a team of six or more engineers.
    What's in the pack

    Toolkit

    Daily carry

    Python, TypeScript, PyTorch, Hugging Face, FastAPI, React, Postgres, Docker, Linux, Git, Claude Code and Codex.

    For the long climbs

    LoRA / PEFT, RAG and retrieval, sentence transformers, continual learning, LLM evaluation harnesses.

    Old boots, still good

    C++, C#, .NET Core, Angular, WPF. Seven years' worth, and it still shows up in how I build.

    Off the trail

    Guitar and piano, table tennis, chess, running, hiking, a garden that mostly survives me.

    Say hello

    Happy to talk about research collaborations, engineering problems that need a model in them, or anything in the projects above.

    vishalbanwari26@outlook.com