Five real projects in the same shape: context → problem → what was built → outcome → what did not work. No client names, and no number I cannot stand behind.
1. Hybrid WAF (2025 — graduation project at production standards)
- Context
- A web environment needing protection that understands context, not only rule matching.
- Problem
- Classic defences rely on fixed signatures: either weak pass-through or blocking that hits real users.
- Built
- Seven filtering layers ordered by cost, an LLM arbitration layer invoked only on doubt, and a hard timeout with a fallback path.
- Outcome
- Contextual decisions at no cost to the healthy majority of requests; under 5% ever reach the arbitration layer.
- What failed
- Classifying every request with a language model was dropped: cost and latency outweighed the benefit. Ordering by cost is what worked.
Full technical case study →
2. DDoS mitigation (2022 — contracted work)
- Context
- A service hit by repeated traffic attacks separated by quiet periods.
- Problem
- The fixed threshold became a problem of its own: legitimate peak traffic resembles an attack in shape.
- Built
- A behavioural baseline learned from normal traffic, early anomaly detection, behavioural fingerprints, and a graduated response instead of instant blocking.
- Outcome
- Legitimate peaks are distinguished from attacks by behaviour rather than volume alone, with legitimate traffic staying online.
- What failed
- Using request rate as the only signal produced more false positives than the service tolerated. Timing and path-diversity dimensions were added.
Full technical case study →
3. Digital reputation protection (2022 — contracted work)
- Context
- A need for early awareness of abusive or copied content appearing online.
- Problem
- Manual searching is slow, and by the time content is found it has already spread.
- Built
- Recurring automated scanning with comparison and matching, instant alerts when matching content appears, and periodic reports.
- Outcome
- Discovery moves from "when someone notices" to "when it appears", with a clear reporting path.
- What failed
- Exact matching alone produced false positives and missed edited copies. Fuzzy matching was added with heavier weight on stable signals.
4. Sodfa Chat (2026 — open source)
- Context
- A personal project testing one hypothesis: can random chat exist with no central storage of the conversation?
- Problem
- Every classic random-chat platform stores conversations and metadata on a central server.
- Built
- Direct pairing between users over a P2P data channel, a transport-and-pairing broker that retains no content, and a fallback path for networks that block peer-to-peer traffic.
- Outcome
- Architectural privacy rather than promised privacy: there is no place to read the content from, because it is never stored.
- What failed
- Direct connection alone failed behind restricted networks. A fallback path was added instead of reporting "offline".
Project page →
5. Digital evidence support for extortion victims
- Context
- People facing extortion or threats who need an organised evidence file, not general advice.
- Problem
- Original content is deleted quickly, and once gone the claim stands unsupported.
- Built
- A public-evidence collection method: recovering archived content, ordering it chronologically, preserving the original as-is with metadata, and producing a file fit for a formal report.
- Outcome
- Preparing the documentation file drops from hours to minutes, and the file becomes submittable.
- What failed
- Relying on screenshots: they do not hold up and carry no metadata. The untouched preserved original is what holds.
Testimonials
I do not publish a credential I have not earned, or a testimonial I wrote myself. This section is ready for two real ones — the LinkedIn profile accepts formal recommendations, and a two-line testimonial from someone who actually worked with you is worth more here than ten later.
Reserved for testimonial 1Client, mentor or colleague — with their consent, in their words.
Reserved for testimonial 2Or a testimonial from cybercrime-victim support work, anonymised.