On a quiet block of the Presidio, at 1004A O'Reilly Avenue, a small San Francisco firm called Ciridae is making a wager that has become unfashionable to say out loud in 2025: that the hard part of enterprise AI is not the model, it is the wiring [Ciridae Privacy Policy]. Founded in 2020 and led by former Andreessen Horowitz gaming partner Jack Soslow, Ciridae sells what it calls AI transformation, a packaged engagement that takes a customer from strategy through data modernization to deployed agents running in production [Ciridae blog].
The pitch is straightforward. Ciridae's services page describes engineers who design and build custom AI tools tailored to a client's business, automating high-friction workflows [Ciridae Services]. The company's framing in its launch post is more ambitious: it talks about a transformation economy where competitive advantage is continuously reborn, and positions itself as the firm that architects those metamorphoses end to end, from strategy to roadmap, data modernization to agent deployment, workflow redesign to change management [Ciridae blog]. In plainer terms, Ciridae wants to be the build partner a Fortune 500 calls when its internal AI council has produced a deck but no shipped systems.
The bet
There is a reason that wedge exists. Most large enterprises in 2025 are sitting on a backlog of AI proofs-of-concept that never made it past a sandbox. The Ciridae founding note is candid that the company was started after years of building AI systems inside large, complex organizations where ambition often hit red tape, and great ideas got lost in PowerPoints [Ciridae About]. That sentence is the entire commercial thesis. If a client has tried and failed to ship internally, an outside team that owns strategy, build, and operate is a cleaner line item than another platform license.
Ciridae's published case work hints at what those engagements look like in practice. One blog post documents a restoration rebuttal automation project, the kind of narrow, document-heavy workflow where an agent can demonstrably move a metric [Ciridae blog]. Another, titled The Future of Coding Is Orchestration, sketches the firm's view that the differentiation now sits above the model layer, in how tools, data, and human review are composed [Ciridae blog]. A testimonial on the services page claims that in just one month, Ciridae understood the client's business, identified the biggest generative AI opportunities and risks, and gave a clear path toward becoming an AI-first company [Ciridae Services]. The customer is not named, but the cadence (one-month diagnostic, then build) matches the standard high-end consulting motion.
Why it could be big
The market shape favors this approach right now. Accenture and Deloitte have publicly booked billions in generative AI bookings, and the boutique tier below them, firms small enough to be technically opinionated but large enough to staff a real engagement, is where a lot of the actual implementation work is landing. A founder-led shop with credible AI taste and a network into the model labs can charge consulting rates while building reusable internal tooling, which is a healthier business than pure staff augmentation.
Ciridae's network is the asset most likely to compress its sales cycle. Soslow spent his a16z tenure on the Games team, where he closed 36 deals in three years and contributed to investments including Yellow [VentureLab UPenn, April 2024] [a16z, May 9 2024]. That is not enterprise software distribution, but it is a Rolodex of operators, a peer group inside one of the most active AI investors in the market, and a credibility marker when a CIO is deciding whether to take the meeting. His X profile, which lists him as CIRIDAE Ambitionist, suggests the firm is being run as a primary bet rather than a side project [X, Jack Soslow].
The team
Co-founder Jack Weissenberger serves as CTO, according to his LinkedIn, and maintains a Hugging Face profile under the handle jw-ciridae, which is at least a public signal of hands-on model work [LinkedIn, Jack Weissenberger] [Hugging Face]. The broader team surfaced through public profiles includes Isabel Moranta, Alex Avendano, Rich Brown, Sumer Mavi, Matthew Donne, and Stas Bondar [LinkedIn, Isabel Moranta]. Bondar and Moranta both have Awwwards-recognized design work tied to Ciridae projects, which tracks with a firm that treats client-facing polish as part of the deliverable [Awwwards, Stas Bondar] [Awwwards, Isabel Moranta]. UK records show a related entity, CIRIDAE LTD, incorporated December 9, 2020 [Tracxn], suggesting the firm has at least contemplated cross-Atlantic delivery.
The honest counterfactual
The bear case on a services-led AI firm is well-rehearsed: margins compress as foundation-model providers, hyperscalers, and the Big Four consultancies all push down into the same workflow-automation engagements, and a boutique without proprietary IP gets squeezed. Ciridae's answer, visible in its own writing, is that the durable wedge is orchestration and operating the systems in production, not just building them [Ciridae blog]. If the firm can convert one-month diagnostics into multi-year operate contracts, the margin profile starts to look more like managed services than like project consulting, which is a meaningfully different business. Whether that conversion is happening at scale is the question the next year of public evidence will answer.
What to watch
The milestones to track are concrete. A first named enterprise customer would validate that the Soslow network converts at the CIO level. A seed round with a recognizable lead would confirm investor appetite for AI-services equity stories, a category that venture has historically been allergic to. And the cadence of the Ciridae blog, which has so far been a useful window into what the team is actually building, is the cheapest signal a reader has on whether the engineering bench is growing.
Technical breakdown
What Ciridae appears to sell, stripped of the marketing layer, is a three-stage engagement: (1) a discovery sprint that maps high-friction workflows and ranks them by automatable value, (2) a custom build phase where engineers compose models, retrieval, tool use, and human-review steps into a production agent, and (3) an operate phase where Ciridae runs the system, monitors it, and iterates. The orchestration emphasis in their public writing implies a stack that treats the LLM as one component among many, with the moat in evaluation harnesses, data pipelines, and the human-in-the-loop design rather than in any single model choice.
What could go wrong at scale
The failure mode for this category is not dramatic, it is gradual. Services firms that win on founder taste tend to hit a ceiling around 30 to 50 engineers, where the senior bench gets stretched thin across too many concurrent engagements and quality drifts. If Ciridae grows past that without either productizing a piece of the stack or building a repeatable delivery methodology that junior engineers can execute, the same red tape and lost-in-PowerPoints failure mode the founders left their previous employers to escape will reappear inside their own org chart. The firms that survive this transition usually do it by extracting one piece of internal tooling and turning it into a product. Whether Ciridae has that piece, and whether it chooses to ship it, is the structural question underneath all the others.