# work.board

Work AI agents, research systems & ML by Rishabh Sharma

20 projects. The top five run live below — scroll and they play. Everything else is further down the board.

# 01 · engram / terminal

Agent memory · Paper

Engram

Does agent memory need LLM-extracted facts — or just the right raw turns?

77.0% vs 77.5% on 778 held-out LoCoMo questions — non-inferior under a pre-registered −5pt margin
Raw turns are 3,061× cheaper to write than LLM-extracted memory
Reranking adds +17.4 pts (LoCoMo) and +9.1 pts (LongMemEval) at k=3

Open case study →

rishabh — ~/engram — zsh — 112×30⌘1

zshpython3.12+

Last login: Thu Oct 15 09:12:44 on ttys004

rishabh~/engrammain ✓make reproduce-v3

▸ stripping API keys from env ................. ok

▸ fetching LoCoMo @ pinned revision .......... ok

▸ fetching LongMemEval @ pinned revision ..... ok

▸ rebuilding numbers.json from committed results

100% tables · figures · pdf

systemktokens/qLoCoMo

turns + jev626577.0 %

engram v2325177.5 %

engram v220—82.4 %

✓ H1 non-inferior Δ −0.5 pt · one-sided 95% bound −3.0 > margin −5.0

✓ write cost 3,061× cheaper than LLM extraction

✓ check.py every number traces to a result file · 0 API calls

✓ built paper_v3/when-does-selection-replace-extraction.pdf

rishabh~/engrammain ✓

zero API calls — every number re-derived

# 02 · engram / the paper

arXiv 2609.34227 · first author

The paper

A pre-registered test of agent memory — and the sentence I'm proudest of.

Listed on Hugging Face Daily Papers
First independent citation within weeks
Pre-registration DOI + Zenodo release

Read on arXiv ↗

arXiv › cs › arXiv:2609.34227 [cs.CL] · v3 · Oct 2026

When Does Selection Replace Extraction? A Pre-Registered Test of Agent Memory with a Typed Decision Model

Rishabh Sharma · Rishika Lall

Abstract

Does conversational memory for agents need LLM-extracted facts, or is selecting the right raw turns enough? We test this under a pre-registered plan on held-out LoCoMo conversations and on LongMemEval. Raw turns selected by one call to a typed decision model are non-inferior to LLM-extracted memory at matched context, while being 3,061× cheaper to write. The context budget decides the answer: when the reader sees few items, selection holds its own; when it sees many, extraction is more accurate.

1 Introduction

Figure 1. Accuracy on 778 held-out LoCoMo questions. Selection matches extraction at small budgets.

PRE-REGISTERED

🤗 Daily Papers · Oct 6cited by Xu & Chen, 2026DOI 10.5281/zenodo.22985242

the sentence I'm proudest of

# 03 · archimedes / studio

Autonomous research · Open source

Archimedes

Give it a prompt. It runs the research loop and hands you a trace you can verify.

Three modes: orchestrated novelty, simple replication, and AlphaEvolve-style evolve
A graph of specialised agents: ideation → novelty gate → plan → implement → check → reflect → write
Novelty scored against the citation graph via arXiv and Semantic Scholar

Open case study →

◆ archimedesmode: orchestrated● run 0412 · live

input

Prompt

sparse MoE routing, low-resource

agent

Ideation

hypotheses

tool

Novelty gate

arXiv · Semantic Scholar

agent

Planner

staged experiments

tool

Implement

Claude Code · sandbox

gate

Success check

criteria vs metrics

agent

Reflector

next stage / revise

output

Writer

LaTeX manuscript

it keeps its failures

# 04 · omniharness / sandbox

Agent harness · Open source

OmniHarness

A self-extending agent platform that builds, registers and verifies its own capabilities.

Long-horizon tasks that take minutes to hours, inside per-thread Docker sandboxes
Self-verifying live preview: the agent can't declare victory while the build is broken
Subagents, persistent memory, hot-loaded skills, MCP client with OAuth

Open case study →

OmniHarnessthread #42 · chai-landingsandbox · docker

🔒 localhost:5173gate: passing

starting dev server…

[plugin:vite:import-analysis] Failed to resolve import "./Hero" from "src/App.tsx". Does the file exist?

/app/src/App.tsx:2:17
1 | import Nav from "./Nav";
2 | import Hero from "./Hero";
  |                  ^
3 | import Menu from "./Menu";

Click outside, press Esc key, or fix the code to dismiss.

Since 2019 · Fishtown

Chai, brewed slow.

Masala, cardamom & ginger — simmered with whole spices. Open 7am till late.

See the menuFind us →

can't say “done” while the build is red

# 05 · area-o1 / dashboard

Local-first · Evidence engine

Area O1

An O-1A / EB-1A evidence file, built on your own computer — every claim traced to its source.

Evidence inbox pulls from GitHub, Hugging Face, Semantic Scholar, OpenAlex, arXiv and ORCID
Every fact is a claim quoting its source word for word
Runs on 127.0.0.1 — no account, no server, no telemetry

Open case study →

Profile · O-1A

Criteria scoreboard

3/8

Original contributions3 evidencebanked
Scholarly articles2 evidencebanked
Judging the work of others1 evidencebuilding
Critical role1 evidencebuilding
Press about the work0 evidencegap
Awards0 evidencegap

omni

an invitation never counts as done

# 06 · rallion / war room

GTM engine · Human-in-the-loop

Rallion

Live market signals in. Evidence-cited, human-approved outreach out.

Signals → evidence-cited opportunities → human approval → end-to-end outreach
Prospects qualified by an active-learning model: GP classifier with BALD acquisition
Every prospect × channel is a thread with a guarded state machine — illegal jumps raise

Open case study →

rallionWar roomThreadsEngine3 new signals

Signals 3

fundingSeries B announcedfintech · 42 employees · 2h

hiring4 data-engineering roleslogistics · 1d

launchPublic API v2 shippeddevtools · 3d

Opportunities 1

Data-platform modernisation3 prospects qualified · GP + BALD

evidence ↗evidence ↗

Approved 1

Data-platform modernisationemail first · paced 9–5 · caps on

Thread state

discovered

qualifying

qualified

contacted

engaged

suppressed · from any state

Approve campaign?

Data-platform modernisation

3 prospects · evidence cited
Email first, LinkedIn optional
Activity window 9–5 · per-channel caps

CancelApprove ✓

rishabh

nothing sends without me

# more.frame

Everything else on the board

Agents & research 5

CAMEL-AI · 2025

DealFlow ↗

A framework for AI sales agents that hold natural, multi-turn sales conversations.

Topology · TTA · 2026

TAMC ↗

Persistent homology as a control signal for time-series test-time adaptation.

CrewAI · FastAPI · 2025

ResArch / Spinscribe ↗

Multi-agent content API with Cognito auth and S3 storage.

OpenAI Agents · 2025

ResumeForgeAI ↗

Tailors a resume to a job description — and to the ATS that reads it.

Agents · 2025

WebWiseAgent ↗

A personal research assistant that works the web for you.

Machine learning 7

GNN · imbalance · 2025

AML P2P Fraud ↗

Caught my own leak: AUPRC 0.988 → an honest 0.380. 26.1% SAR conversion at 2% alert capacity.

LightGBM · CI/CD · 2025

Home Credit Default Risk ↗

Credit-risk modelling for the unbanked, shipped behind a Jenkins pipeline.

Legal ML · 2022

Sentencing Prediction ↗

Predicting sentencing type from court docket data, used at Drexel's Kline School of Law.

Computer vision · 2023

Wildfire Perimeters ↗

Wildfire and perimeter detection from GeoTIFF imagery.

XGBoost · CatBoost · 2025

Collision Severity ↗

Predicting injury severity from crash data with gradient-boosted ensembles.

Deep learning · 2025

Customer Segmentation ↗

Estimating family size from customer attributes and purchase patterns.

Classification · 2025

NBFC Loan Foreclosure ↗

Predicting early loan foreclosure for a non-bank lender.

Data engineering 3

ETL · Redshift · 2025

GoodReads Pipeline ↗

Data pipeline for GoodReads API data into a Redshift warehouse.

Airflow · Spark · 2025

Airflow + EMR ETL ↗

Airflow-orchestrated Spark jobs on EMR, from S3 back to S3.

Time series · 2023

Covid-19 Forecasting ↗

Forecasting cases and deaths in Mexico, with a dashboard.