Chadsel LLC
Custom AI Bots.
Research at Fleet Scale.
We build bot fleets for investment research: idea generation, equity and credit work-ups, trading signal generation and backtesting, and the middle- and back-office work behind them. Built for your mandate, your universe, and your data — not configured out of a product.
One constraint runs through every deployment: MNPI, positions, and trade data never transit a third-party API. On-premise inference for the restricted side, isolated cloud for public-data research, and a decision that stays with your PM.
What We Build
Bots with a job description.
A fleet is not one assistant with a bigger prompt. Each bot has a defined input, a defined output, and an owner — which is also what makes the output reviewable.
Screener bots
Run continuously across filings, transcripts, news wires, market data, and alternative data to surface names that fit your mandate — long or short — and reject the rest with a written reason you can audit.
Research bots
Take a ticker and build the work-up: business model, unit economics, competitive position, capital structure, management history, and the bear case stated as carefully as the bull case, with every claim linked to the source document.
Signal bots
Formulate a rule-based hypothesis, write the code, pull the data, run the backtest, and report the result with the null results included — including the ones that kill the idea.
Document bots
10-K and 10-Q, credit agreements, indentures, side letters, DDQs, and prime broker statements — parsed, cross-referenced, and searchable, so a question about a covenant gets answered out of the covenant.
Desk bots
Middle and back office: reconciliation breaks, position and cash tie-outs, trade file normalization across prime brokers and the fund administrator, NAV package checks — the recurring work that consumes an operations team.
Orchestration
A supervising agent that routes work between bots, enforces the data classification boundary, retries failures, ranks output for human review, and records which bot touched which data and when.
Front, Middle, Back
We have run all three, not just the front one.
Our team built and ran front-, middle-, and back-office technology at Christofferson Robb and Company, LLC, a $9B structured credit fund across New York and London — trade capture through reconciliation, administrator and prime broker files, NAV packages, and the research stack above it.
That matters here for one reason: a research bot is only as good as the data plumbing under it, and most of that plumbing lives in the middle and back office. Firms that skip it end up with a bot that writes confident analysis off stale positions.
Front office
Workflows
Idea generation, screening, fundamental and credit work-ups, PM support, market and alternative data
Where bots fit
Coverage and first drafts. A bot fleet reads everything and hands the PM a ranked shortlist with its reasoning attached.
Middle office
Workflows
Trade capture and allocation, P&L and exposure reporting, risk limits, counterparty and margin monitoring
Where bots fit
Exception handling. Bots watch for breaks, limit breaches, and stale marks, and escalate with the evidence assembled.
Back office
Workflows
Reconciliation, cash and position tie-outs, prime broker and administrator files, NAV packages, treasury and fees
Where bots fit
The recurring grind. Bots normalize files, reconcile, and draft the break commentary an operations analyst would write.
Where the Market Is
The demand is in research, not execution.
The programs that made news through 2025 and 2026 all point the same way: bots are being put on idea generation, company analysis, and signal discovery — the analyst bench — while execution and risk automation stay comparatively quiet. And in every one of them, a human still owns the decision.
Named below as public reference points, with sources. They are not clients of Chadsel LLC and nothing here implies an endorsement.
Magnetar Capital
$18B AUMA new fund built to run without human analysts: hundreds of AI bots search for ideas, study companies, recommend positions, and forecast trends. Humans still make the final call on every trade.
Man Group — AlphaGPT
Man NumericAgentic AI that mines data, formulates rule-based signals, writes the code, and backtests them before a human sees the output. Reported to produce research concepts in minutes rather than days, with several dozen signals approved for live trading.
Abundance
$100M seedThousands of bots on idea search, deep research, stock selection, position sizing, and execution, maintained by a small team of quantitative researchers and engineers instead of a fundamental PM bench.
Deployment Architecture
Three models. One requirement: MNPI stays yours.
Every deployment starts with data classification. We recommend the architecture after understanding what your bots will read — not before.
On-Premise
- Data location
- Your infrastructure
- MNPI exposure
- Safe
- Compliance posture
- Maximum
- Best for
- MNPI, positions, restricted lists
Isolated Cloud
- Data location
- Dedicated Azure or AWS tenant
- MNPI exposure
- Configurable
- Compliance posture
- Strong
- Best for
- Public-data research at scale
Hybrid
- Data location
- Split by classification
- MNPI exposure
- Routed on-premise
- Compliance posture
- Configurable
- Best for
- Most funds running bot fleets
Platforms
What the bots are built on.
The bot is the product; these are the parts. Chosen for institutional workflows, data security requirements, and the compliance constraints of investment management.
OpenClaw
Open-Source IntegrationOur agent integration layer for capital markets, and the piece that makes a fleet operable: it connects bots to Bloomberg, FactSet, OMS/PMS, prime broker, and administrator data, holds the classification boundary, and keeps a record of which bot read what. Deployable on-premise or in an isolated Azure or AWS tenant — no trade data leaves your infrastructure.
Claude (Anthropic)
Reasoning model behind research and document bots: filings analysis, credit agreement review, LP report drafting, regulatory Q&A. Deployed in an isolated tenant — no data shared across tenants.
Codex / Computer-Use
Writes and runs the code a signal bot needs, and drives the portals that have no API — prime broker sites, administrator uploads, internal tools built before anyone thought about integration.
Open Harness
Agent orchestration with data classification routing — MNPI-flagged and restricted-list data stays on-premise; general research queries route to cloud models with an audit trail.
Ollama / vLLM
On-premise LLM inference for the restricted side of the boundary. Runs Llama, Mistral, and Phi models on your hardware, with zero egress for MNPI, position data, or anything subject to Regulation S-P.
LangChain
The plumbing under a bot fleet: tool use, retries, rate control, and connections to market data feeds, internal research databases, and custodian APIs.
LlamaIndex
Indexes internal research, filings, DDQ responses, and LP documents so retrieval is grounded in your own corpus rather than a model's memory — fully on-premise.
Investment
Custom bot development pricing.
Engagements scope to your mandate, data classification, and compliance requirements. Most funds start with one bot role on one workflow before expanding to a fleet.
| Engagement | Starting From | Typical Range |
|---|---|---|
| AI readiness & governance assessment | $5,000 | $5,000–$12,000 |
| Private AI / agent pilot | $20,000 | $20,000–$60,000 |
| Institutional production deployment | $75,000 | $75,000–$250,000+ |
| Specialist consulting | $175/hr | $175–$275/hr by labor category |
| Managed AI operations | $2,000/mo | $2,000–$6,000/mo |
Ranges exclude third-party licenses, hardware, cloud consumption, model usage, market data, and compliance testing unless listed in the signed scope.
Infrastructure Stack
What a fleet runs on.
Air-gapped where it has to be, auditable everywhere, and sized against your workflows rather than a vendor reference architecture.
| Layer | Technology |
|---|---|
| Local LLM Runtime | Ollama, vLLM (air-gapped) |
| API Proxy | LiteLLM (OpenAI-compatible) |
| Bot Orchestration | LangChain, Open Harness |
| Research Inputs | Filings, transcripts, news, market and alternative data connectors |
| Document Intelligence | LlamaIndex, pgvector, Qdrant |
| Signal Evaluation | Backtest harness with reserved hold-out windows, transaction-cost and capacity models |
| Fund Systems | OpenClaw (Bloomberg, FactSet, OMS/PMS, prime broker, administrator) |
| Cloud AI Isolated | Azure AI Foundry (dedicated tenant), AWS Bedrock |
| Containerization | Docker, Kubernetes, KServe |
| Data Security | Encryption at rest/transit, RBAC, per-bot data scopes, immutable audit trail |
| Open Models | Llama 3, Mistral, Phi-3 (on-prem); GPT-4o, Claude (isolated cloud) |
Controls
Built for auditors, not just for the desk.
Autonomy raises the evidentiary bar rather than lowering it. Every fleet we build ships with the controls an examiner would look for: audit trails, access logs, data residency documentation, provenance on generated output, and a written human decision point.
We have built and operated cybersecurity programs that passed both SEC and FCA examinations. That background shapes how these architectures get designed — defensively, with documentation that holds up under scrutiny.
Human-in-the-loop by design
Bots produce coverage, analysis, and candidate signals. The investment decision stays with your PM or investment committee, and the handoff point is written into the scope before the first bot ships.
Provenance on every output
Each claim carries the document, page, and timestamp it came from. An output nobody can trace back is a liability, not research.
Rule 17a-4
Immutable audit log architecture for bot-generated outputs subject to books-and-records obligations.
Regulation S-P
Data handling and access controls reviewed against SEC Regulation S-P requirements before deployment.
NIST SP 800-53
Zero-trust access, encryption at rest and in transit, and incident response aligned to NIST 800-53.
FCA SYSC
Operational resilience and data governance architecture reviewed against FCA SYSC 8 requirements for UK-regulated entities.
Common Questions
Frequently asked.
Can AI bots actually replace an analyst team?
They replace the coverage and the first draft, not the decision. A bot fleet can read every filing, transcript, and news item in your universe every day, which no analyst bench can do — and the publicly reported programs doing this at scale, including Magnetar and Abundance, still keep a human on the final trade decision. What we build is that shape: bots generate, screen, analyze, and document; your PM or investment committee decides, and the handoff point is written into the scope before anything ships.
How do you keep a signal bot from producing overfit backtests?
The same way a research committee does it, enforced in code: hypotheses are recorded before the test runs, hold-out periods are reserved and not reused, failed and null results are logged rather than discarded, and transaction cost, turnover, and capacity assumptions are stated in every result. A bot that can run ten thousand backtests will find something in noise — so the constraint is not throughput, it is the evidence standard the output has to meet before a human sees it.
How does MNPI stay contained when a fleet of bots is reading everything?
By classifying data before the bots touch it and routing on that classification. Restricted and MNPI-flagged data is processed by on-premise inference — Ollama or vLLM on hardware you control — and never transits a third-party API. Each bot runs with an explicit data scope rather than blanket access, egress is logged, and the architecture is reviewed against SEC Regulation S-P and FCA SYSC data handling requirements before go-live.
What data do the bots need, and what if our data is a mess?
Public-side bots need filings, transcripts, news, and market data, all of which we connect. Anything that touches your positions, P&L, or reconciliations needs your own systems — OMS/PMS, prime broker and administrator files, the internal research store. That side is usually where the real work is, and it is the part we have done before: our team spent 13 years running front-, middle-, and back-office technology at Christofferson Robb and Company, LLC, a $9B structured credit fund across New York and London. The clean data layer comes before the bots, not after.
Do we need our own GPUs to run this?
Only for the restricted side. On-premise inference for MNPI-sensitive work needs hardware you control, and for most funds that is one or two GPU servers, not a cluster. Public-data research can run in a dedicated cloud tenant with no cross-tenant data sharing. The readiness assessment sizes this against your actual workflows rather than a vendor reference architecture.
New to Chadsel LLC?
Start with one bot role, on one workflow.
A $5,000–$12,000 readiness and governance assessment defines the target workflow, the MNPI boundary, the human decision point, the deployment architecture, and the acceptance criteria the first bot has to meet.
(212) 749-9408