Investment OS for private-company screening.
The investment team was drowning in the first pass: normalizing exports, checking fit, looking up registry context, writing the same notes, rebuilding a company's history on every new screen. We built a system that runs that pass for them. It takes the whole list, tens of thousands of companies in a single run, and deterministic filters drop them before the model spends a token. Every relevant company goes through a bounded research loop: the agent pulls KRS registry data, financials, LinkedIn signals, and web context, answers a fixed 11-question framework, and writes a score with thesis and risks. The dossier stays in memory, so a company that returns carries its history. The agent prepares the memo, the contacts, and the outreach draft, but the analyst approves before anything leaves the building. The public screenshots are real captures from the running system, with company and account data redacted.
For whom
Use this pattern when analysts need first-pass investment judgment across a private-company universe, with sources, memory, and override controls, not another chatbot summary.
Upload → gates → dossier → memo
- 01Hard filters drop companies before the model spends a token
- 02Every relevant company has its own sourced dossier
- 03Weak evidence lowers confidence instead of sounding sure
- Sector
- Private-company screening (PE/VC)
- Inputs
- Company exports, KRS registry, financials, LinkedIn, web context
- Architecture
- Deterministic gates before AI plus a bounded agent loop
- Surfaces
- Web and Telegram chat (one agent), Slack notifications
- Memory
- Company history across runs, analyst overrides, deal stage
- Status
- Delivered
Operating loop
Gates first. Analysis after.
The system cleans the input, applies the team's criteria as code, and only then sends relevant companies into deeper analysis. The model never sees the companies that fail hard rules.
Company list upload and validation
Deterministic gates: financials, scale, ownership
Enrichment from sources: KRS, financials, LinkedIn, web
Agent analysis: 11 questions, score, thesis, risks
Durable company memory and dossier across runs
Memo, C-level contacts, and outreach draft
Analyst approval before action
Architecture
A screening system, not a prompt.
The useful part is the control plane around the model: deterministic gates before analysis, tools inside the loop, durable memory after each run, and analyst approval before business action.
- 01
Ingest
Normalize the company universe
Excel, CSV, and TSV exports can arrive with inconsistent headers, mixed languages, missing identifiers, and stale financials. The system maps the input into a screening schema and drops companies already scored in earlier runs.
- 02
Gate
Run deterministic checks first
Revenue and growth thresholds, margin, debt-to-equity, headcount scale, ownership, and analyst criteria are applied as code. Companies that fail hard rules are documented as such; AI is reserved for companies worth deeper research.
- 03
Research
Use tools, keep the evidence chain
In a bounded loop the agent pulls the companies to analyze and a fixed investment framework, enriches each one with KRS registry data, financials, headcount history and hiring signals from LinkedIn, and web context, then writes the thesis with source references, missing-data flags, and confidence language.
- 04
Remember
Carry findings across screens
Company context, prior scores, qualitative answers, known unknowns, analyst overrides, and deal stage are stored on one card, so repeated companies are compared against their own history instead of starting cold.
Why this is agentic
A chatbot summarizes the file you gave it. Investment OS runs a bounded workflow with state, tools, memory, scoring, exception handling, and human approval.
- Deterministic gates before AI
- Tool-backed evidence collection
- Long memory per company
- Review, override, and audit trail
Operating proof
What makes it hard to copy.
The hard part is the control plane: hard gates before AI, research across four signal classes (registry, financials, LinkedIn, web context), memory between runs, and the analyst holding the last word. The value is in what enters, what gets stopped, and what is sourced.
- Work
AI starts after structured screening, not before it.
The first pass is deterministic: input mapping, financial checks, scale filters, ownership signals, and analyst criteria. That reduces cost to a couple of cents per scored company and keeps the model away from companies that fail hard rules.
- Escalation
Weak inputs become flags, not false certainty.
Missing identifiers, conflicting names, stale financials, low-confidence context, and hard-rule failures are surfaced as review conditions instead of being smoothed into confident text.
- Measurement
The run is inspectable while it works, not just at the end.
Operators see stage gates, pass rates, verdict buckets, scored and unscored companies, and export state. A large run, on the order of two hours per 10,000 companies, is visible as it goes, so screening coverage is never hidden in the background.
- Trace
Every score can be reconstructed and defended.
The score, the verdict, each analyst override, and every stage change are saved with source and timestamp. The team returns to the evidence the verdict stood on instead of trusting a number with no history.
Agent work
The agent compresses first-pass screening, but leaves the investment decision with the analyst.
Problem
Before an analyst even began judging a company, the time went on tedious prep work: tidying up messy exports, checking whether a company fit at all, digging up context in registries. The same first-pass notes were written from scratch, and when the same company came back in a later screen, its history had to be rebuilt by hand.
How it works
The system takes on the entire first pass, even across tens of thousands of companies at once. First, hard rules (financials, scale, ownership structure) filter out what would be rejected by hand anyway, so only companies worth a closer look move on. The agent studies each one like a first-pass analyst: it pulls KRS registry data, financials and LinkedIn signals, answers 11 diligence questions with sources cited, assigns a score, and writes a thesis along with the risks.
Memory across runs
Each company has one card in the knowledge base, no matter how many times it returns. The system keeps score history, prior verdicts, qualitative answers, known unknowns, analyst notes and overrides, and the stage the deal sits at. When a company comes back in a later run, that context is injected into the analysis, so the agent compares it against its own history instead of scoring it cold.
Outreach and follow-up
For companies with a go verdict, the agent pulls C-level contacts and prepares a contact draft: email or LinkedIn message, in Polish or English, in a chosen tone, deliberately without mentioning the scoring or the AI analysis. Drafts wait in the deal pipeline, from research through outreach to meeting and diligence, and each conversation's status is updated by hand. A watchlist monitors selected companies and shows what changed since the last snapshot.
Boundaries
The system prepares the work, but it does not make the investment decision, override the analyst, or send any outreach on its own. Draft emails and messages wait for sign-off; nothing goes out automatically. Every company keeps sources, data gaps, score history, analyst overrides, and a screening trail, so the team can inspect why the verdict exists.
What broke
The hard part was not making the agent write. It was making it stay honest when the data was uneven. Raw exports had missing fields, synthetic identifiers, stale financials, and company names that did not match across sources. We added validation, explicit gap flags, source retention, name normalization to match registry against export, and unknown states, so weak inputs stay visible instead of becoming confident prose.
Result
The analyst starts from a ranked pipeline and a sourced company dossier, not from raw rows. A screening batch produces stage-level pass rates, verdict buckets, per-company financial context, AI assessment, review controls, and a source trail. Repeated companies carry memory from previous runs, and the memo and outreach draft wait ready for approval, so the analyst's time goes to the decision, not to rebuilding context from scratch.
Work surfaces
The pipeline is the control room. The dossier is the memory.
The analyst starts from a ranked pipeline and a sourced company dossier, not from raw rows. A screening batch produces stage-level pass rates, verdict buckets, per-company financial context, AI assessment, review controls, and a source trail. Repeated companies carry memory from previous runs, and the memo and outreach draft wait ready for approval, so the analyst's time goes to the decision, not to rebuilding context from scratch.
Run
The control room for the whole batch: where companies stand and what cleared the gates.
Dossier
One company's memory: verdict, thesis, and risks next to what backs them.
Analysis
11 diligence questions across four lenses (business model, market, growth readiness, red flags), turned into a reviewable record with scores and confidence.
Real redacted screenshots
The product surface is the proof.
The product screenshots come from the system running in Docker. The operator-chat screenshots come from a real mobile work surface. Client and account data, chat workspace identity, and exact batch labels are redacted for public use.
Operator chat on a phone
The analyst runs screening from chat.
We framed three Telegram captures in a phone and redacted company and account data. The analyst asks for a shortlist, checks the product, and expands one company without leaving the chat.
01The analyst asks for a top-10 shortlist and receives a ranked first-pass list for deeper review. 
02The agent checks whether product names are real productized solutions or just marketing wrappers. 
03One company expanded with registry context, revenue signals, thesis, and caveats before review.
- Screen 01vc-investment
01Run overview: staged gates, verdict distribution, pipeline filters, per-company score, and export control. - Screen 02vc-investment
02Company dossier: score, thesis, risks, financial context, KRS registry data, the 11 diligence questions, analyst notes, and review controls.
Autonomy boundary
Boundaries
The system prepares the work, but it does not make the investment decision, override the analyst, or send any outreach on its own. Draft emails and messages wait for sign-off; nothing goes out automatically. Every company keeps sources, data gaps, score history, analyst overrides, and a screening trail, so the team can inspect why the verdict exists.
The hard part was not making the agent write. It was making it stay honest when the data was uneven. Raw exports had missing fields, synthetic identifiers, stale financials, and company names that did not match across sources. We added validation, explicit gap flags, source retention, name normalization to match registry against export, and unknown states, so weak inputs stay visible instead of becoming confident prose.
Recognize your own process here? Let's see what an agent could take over.
- 30 minutes with the engineer who would build it, not a salesperson.
- A review of the processes that cost you the most time and money.
- A written summary: what to automate, in what order, with cost ranges.
No sales deck and no obligations. If automation doesn't make sense, we'll write that too.